ArXiv: 2512.07584

🎯 Pitch

A compact 6B-parameter diffusion model matches or beats 20B+ alternatives like FLUX.1-dev and Qwen-Image, achieving 90.7% accuracy on Chinese text rendering—nearly 35 points above the prior open-source best—by encoding quoted characters individually to bypass subword token failures. This efficiency leap suggests meticulous data curation and targeted design can replace brute-force scaling, slashing deployment costs while setting a new state-of-the-art.

1. Executive Summary

This technical report introduces LongCat-Image, a bilingual (Chinese-English) image generation foundation model that prioritizes efficient architectural design over brute-force parameter scaling, using a compact 6B-parameter hybrid MM-DiT and Single-DiT diffusion transformer paired with a Qwen2.5VL-7B text encoder. The paper establishes new state-of-the-art performance through a coordinated multi-stage training pipeline spanning progressive pre-training, mid-training, SFT, and RLHF with curated reward models — including an AIGC detection model to guide photorealism — alongside a character-level tokenization strategy for text rendering (encoding quoted text at character granularity rather than subword tokens) and a Multi-Granularity Captioning framework (hierarchical captions from entity tags to dense photographic descriptions). The model achieves 90.7% overall accuracy on the ChineseWord benchmark covering 8,105 characters — dramatically outperforming the next-best open-source model at 56.6% — and delivers strong text-to-image alignment (87% on GenEval, 86.8% on DPG-Bench), while its editing variant sets open-source SOTA on CEdit-Bench, GEdit-Bench, and ImgEdit-Bench, in some cases surpassing commercial systems. The work demonstrates that a 6B model can compete with architectures ~3×–13× larger, establishing that efficient design and rigorous data curation can substitute for parameter count across photorealism, text rendering, and editing tasks — but only when training data is meticulously filtered for AIGC contamination and quality, and only when RL reward models directly target the specific failure modes (AIGC-like textures, OCR errors) that brute-force scaling alone fails to resolve.

2. Context and Motivation

The Core Problem: Parameter Scaling Has Not Delivered Proportional Gains in Image Generation

The fundamental issue this paper confronts is that the dominant strategy in image generation model development — aggressively scaling up model parameters — has reached a point of diminishing returns while introducing escalating costs. The paper's opening salvo in Section 1 is direct:

"a prevailing trend has been the dramatic scaling of model parameters — from PixArt-α at 0.6B, to Stable Diffusion 3.0 at 8B, and further to Qwen-Image at 20B and even larger Mixture-of-Experts (MoE) architectures like Hunyuan-3.0 with 80B full parameters. The expectation has been that, similar to Large Language Models (LLMs), diffusion models would experience a breakthrough in performance through brute-force scaling. However, our observations reveal a critical issue: unrestrained parameter growth has not delivered the anticipated qualitative leap."

This is not a casual observation — it reflects a specific empirical finding by the LongCat team that increasing diffusion model parameters beyond a certain threshold produces severely diminishing returns in visual quality, while simultaneously creating three concrete problems:

1. Soaring computational costs. Training an 80B-parameter MoE model like Hunyuan-3.0 requires enormous compute budgets that are inaccessible to most research labs and startups, effectively concentrating image generation capability in the hands of a few well-resourced organizations.

2. Higher deployment barriers. A 20B model like Qwen-Image requires substantial VRAM for inference — the paper notes that the field has gravitated toward "nearly 20B or larger" architectures (Abstract), which places them beyond the reach of consumer GPUs and makes edge deployment impractical. For context, a 20B-parameter diffusion model might require 40-80GB of VRAM depending on precision and implementation, far exceeding what most developers can access.

3. Increased inference latency. Larger models are slower per generation step, making them unsuitable for interactive applications, real-time editing, or high-throughput production environments. The paper does not provide specific latency comparisons, but the implication is clear: scaling parameters while ignoring efficiency creates models that are powerful on benchmarks but impractical in deployment.

The paper's core thesis, drawn from the LongCat team's design philosophy of "Building efficient and powerful model" (Section 1), is that there exists a more optimal equilibrium between state-of-the-art performance and efficiency — and that this equilibrium sits at roughly 6B parameters, not 20B or 80B.

Why This Problem Matters: The Stakes for Open-Source and Democratization

The gap this paper addresses has implications beyond academic interest. The parameter-scaling trend actively threatens the democratization of image generation technology. When only well-funded commercial labs can train and deploy the best models, the open-source community is left with weaker alternatives, and researchers without massive compute budgets cannot reproduce or build upon leading results. The paper makes this explicit:

"This not only hinders the democratization of the technology but also poses challenges for open academic research." (Section 1)

The commercial consequences are equally significant. The paper's benchmarks — particularly the internal Poster & Scene Text evaluation (Table 8) — are drawn from "business-critical scenarios such as poster design and natural scenes with text" (Section 5.1). These are not abstract academic tasks; they represent real production workloads for the Meituan LongCat team, suggesting that the efficiency-performance tradeoff directly affects commercial viability of image generation products.

The Specific Capability Gaps in Existing Models

Beyond the general efficiency problem, the paper identifies three specific capability gaps that existing models — both open-source and commercial — fail to adequately address:

Gap 1: Chinese Text Rendering Remains Fundamentally Broken

This is perhaps the most concrete and measurable gap the paper identifies. The Chinese character set exhibits an extreme long-tail distribution: approximately 3,000 common characters appear frequently in natural data, but over 5,000 rare characters appear so sparsely that models trained on conventional data distributions simply never learn to render them. The paper quantifies this explicitly in Section 4.1:

"The Chinese character set exhibits a distinct long-tail distribution, comprising approximately 3,000 common characters and over 5,000 rare ones that appear sparsely in natural data."

The consequences are stark in the ChineseWord benchmark results (Table 7). Qwen-Image — the previous open-source leader — achieves only 37.1% accuracy on Level 2 characters and a catastrophic 6.1% on Level 3 (the rarest characters). Even the commercial Seedream 4.0 achieves only 41.2% and 2.3% respectively. These are not marginal failures; they represent near-total collapse on rare characters.

This matters because Chinese text rendering is not a niche requirement. For any image generation model deployed in Chinese-speaking markets (a population of over 1.4 billion), the ability to render names, locations, product labels, signage, and artistic text in Chinese is a core functional requirement, not an optional enhancement. A model that can only render 3,000 characters — roughly the literacy level of an elementary school student — cannot serve professional design, advertising, or publishing use cases.

The underlying cause is architectural: prior models predominantly used CLIP and T5 text encoders (Section 3.2), which were designed for English and lack adequate Chinese tokenization. Even models that adopted LLM-based encoders — like Qwen-Image using Qwen2.5-VL — achieve improved but still severely limited coverage because the subword tokenization used by LLMs creates a combinatorial explosion when encoding rare Chinese characters. Each rare character may be split into multiple subword tokens, making it exponentially harder for the diffusion model to learn the character's visual form.

Gap 2: Photorealism Degradation from AIGC-Contaminated Training Data

The paper makes a striking empirical observation about training data quality that is not widely discussed in the literature:

"We observe that even a small proportion of AIGC-contaminated data can cause the model to prematurely converge to a narrow local optimum during training. While this may accelerate initial convergence, it severely limits the model's potential to achieve higher levels of realism during subsequent fine-tuning." (Section 1)

This is a crucial insight. Most large-scale image datasets scraped from the web now contain a growing fraction of AI-generated images. The paper's claim is that these synthetic images — even in small quantities — act as a kind of quality poison that constrains the model's ultimate photorealism ceiling. The mechanism, as described in Section 2.1.4, is that during early pre-training, exposure to even 0.5% artistic or synthetic data biases the model toward learning simplified visual patterns, "effectively causing a 'collapse' in the realistic generation subspace."

This explains a phenomenon that many practitioners have observed but struggled to diagnose: models trained on web-scale data often produce images with a characteristic "plastic" or "greasy" texture (Section 2.1.1) that distinguishes them from genuine photographs. The paper's diagnosis is that AIGC contamination — not just architectural or training limitations — is a primary cause.

The practical implication is significant: data curation, not just model scaling, determines photorealism. This challenges the implicit assumption of the scaling paradigm that more data (regardless of quality) plus more parameters yields better results.

Gap 3: Visual Consistency in Image Editing Remains Unsolved for Open-Source Models

The paper identifies image editing as a separate but equally challenging domain where existing open-source models fall short. The two primary challenges are defined in Section 1:

"executing editing instructions with high precision and maintaining strict visual consistency between the original and edited images"

Visual consistency — ensuring that unedited regions of an image remain pixel-identical while edited regions change according to instructions — is particularly difficult because it requires the model to simultaneously understand what to change, what to preserve, and how to blend changes seamlessly into the existing visual context. The paper notes that "a significant gap remains in achieving a seamless and reliable editing experience" (Section 1), and that open-source models like FLUX.1 Kontext [Dev] and Step1X-Edit achieve substantially lower scores on editing benchmarks than commercial systems (Tables 9, 10, 11).

What makes this problem hard is that it sits at the intersection of instruction following, visual understanding, and generative quality. A model must not only understand "change the wall color to yellow" but also identify exactly which pixels constitute the wall, preserve the wall's texture and lighting, and ensure the new color integrates naturally with the scene — all without altering the furniture, floor, or people in the room.

Where Prior Approaches Fall Short

The paper identifies several specific shortcomings in prior work, organized by the aspects of the problem they fail to address:

Architectural Shortcomings

Prior models predominantly used CLIP and T5 text encoders (Section 3.2), which have fundamental limitations for non-English languages. CLIP was trained on English image-text pairs and has poor coverage of Chinese characters and concepts. T5, while multilingual, uses subword tokenization that fragments rare characters into multiple tokens, degrading the conditioning signal for text rendering.

Additionally, prior models often injected text embeddings into timestep embeddings for adaLN modulation (referenced in Section 3.2 as a "conventional" approach that the paper abandons), which the authors found empirically to provide negligible performance gains while adding architectural complexity.

In the editing domain, the paper notes that existing approaches (FLUX.1 Kontext, Step1X-Edit) achieve reasonable results but exhibit clear quality degradation compared to commercial alternatives — particularly in maintaining consistency across multi-turn edits where errors compound with each sequential operation.

Data Curation Shortcomings

The paper identifies a critical, widespread failure in data curation: insufficient filtering of AIGC content. Most large-scale training pipelines do not rigorously exclude AI-generated images, and the paper's claim is that this contamination is a primary cause of the "plastic" texture that distinguishes AI-generated images from genuine photographs.

A second data shortcoming is the lack of systematic captioning granularity. Prior work typically used single-style captions from a single VLM, which produced three problems (Section 2.1.3): (1) insufficient integration of world knowledge (entities, brands, proper nouns are often described generically), (2) restricted diversity in caption formats (leading to model brittleness when faced with user prompts that differ from training caption style), and (3) low information density (verbose captions consume token budget without adding useful detail). The paper's Multi-Granularity Captioning framework explicitly addresses all three.

Verification and Alignment Shortcomings

Existing models largely relied on simple aesthetic scoring for data filtering and lacked fine-grained reward signals during RL. The paper argues that single-metric approaches fail to capture the multidimensional nature of image quality. Its Comprehensive Aesthetics Evaluation pipeline (Figure 7) decouples evaluation into Quality (technical fidelity: saturation, contrast, color richness, MUSIQ, Q-Align) and Artistry (photographic merit: composition, lighting, shadow, color tonality, Q-Align-Aesthetics), using an ensemble of six complementary methodologies.

More critically, prior work lacked adversarial signals against specific failure modes. The paper's innovation of incorporating an AIGC detection model as a reward model during RL (Section 1) directly targets the photorealism problem: rather than just rewarding "good" images, it penalizes images that exhibit synthetic artifacts. This adversarial approach — using a discriminator trained to detect AI-generated content to push the generator away from that manifold — is conceptually similar to GAN training but applied at the RL alignment stage of a diffusion model.

Training Strategy Shortcomings

The paper identifies that naïve resolution progression in pre-training — jumping directly from 256px to 1024px — causes computational instability (Section 4.1). Its solution is a three-stage progressive curriculum (256px → 512px → 512-1024px) with an explicit intermediate stage, which is a refinement rather than a radical departure from prior work but reflects empirical observations about training stability.

More substantively, the paper argues that initializing editing models from fully-aligned SFT/RL checkpoints is suboptimal (Section 6.3.1). The reasoning is insightful: SFT and RL narrow the model's state space to a specific aesthetic manifold, which makes the model less plastic and less able to generalize across diverse editing tasks. Initializing from a mid-training checkpoint — which has broader generative capabilities but less aesthetic constraint — provides a better foundation for the model to learn editing while maintaining flexibility. The paper states:

"We deliberately choose to initialize the editing model with weights from the mid-training stage of the T2I model, rather than from a highly optimized state after SFT or RL. The latter models exist in a narrowed state space, which is less conducive to learning and generalizing across diverse editing tasks." (Section 1)

How This Paper Positions Itself

The paper positions LongCat-Image as a counter-argument to the parameter-scaling orthodoxy, but not a rejection of scaling altogether. The 6B parameter count is deliberately chosen — it is not an arbitrary small model but rather what the authors determined through "systematic experimentation" to be "the ideal foundation for balancing capability and efficiency without compromising generative quality" (Section 1). This is an important nuance: the claim is not "smaller is always better" but rather that there exists a Pareto-optimal point where further parameter scaling yields negligible quality gains per unit of compute while incurring significant deployment costs.

The paper's positioning relative to specific competitors is explicit:

  • Against Qwen-Image (20B): LongCat-Image aims to match or exceed its text-to-image alignment and text rendering quality while using less than one-third the parameters. The benchmarks largely support this claim — 87% vs. 87% on GenEval, 95% vs. 93% on GlyphDraw2 average, and 90.7% vs. 56.6% on ChineseWord — though on DPG-Bench, Qwen-Image maintains a lead (88.32% vs. 86.80%).

  • Against HunyuanImage-3.0 (80B MoE): LongCat-Image aims to dramatically outperform a model more than 13× larger. The benchmarks show this clearly across the board: 87% vs. 72% on GenEval, 86.8% vs. 86.1% on DPG-Bench, 95% vs. 78% on GlyphDraw2, and 90.7% vs. 49.3% on ChineseWord.

  • Against Seedream 4.0 (commercial, closed-source): LongCat-Image positions itself as competitive with top-tier commercial systems in specific dimensions. On GenEval, it trails slightly (87% vs. 84%), and on DPG-Bench (86.8% vs. 88.25%). However, on text rendering benchmarks like GlyphDraw2 (95% vs. 97% average) and ChineseWord (90.7% vs. 58.5% — a massive gap on rare characters), it achieves parity or superiority in the Chinese rendering domain specifically. Human evaluation shows LongCat-Image achieving parity in Text-Image Alignment and Visual Plausibility, with a slight edge in Visual Realism over Seedream 4.0, while trailing in Aesthetics (Section 5.2).

The paper's unique value proposition is not that it dominates all benchmarks — it doesn't — but rather that it achieves competitive-or-superior performance at a fraction of the parameter count while providing specific capabilities (Chinese text rendering, photorealism via AIGC-aware training) that larger models lack. The efficiency advantage is not just about model size; it's about deployment feasibility. A 6B model can run on a single consumer GPU, while a 20B or 80B model requires datacenter hardware. For the open-source community that the paper explicitly aims to serve, this is the difference between accessible and inaccessible.

Finally, the paper distinguishes itself through its open-source philosophy. The authors release not just final model weights but also mid-training checkpoints (the "Developer Version" in Section 4.2), the complete training codebase, and the CEdit-Bench evaluation suite. The mid-training checkpoint release is particularly noteworthy: it provides a model that has broad generative capabilities but is not yet locked into a specific aesthetic through SFT/RL, making it suitable as a base for downstream fine-tuning. This is unusual in the field, where most releases are either final aligned models or pre-training checkpoints, without the intermediate stage that the paper argues is optimal for editing task initialization.

3. Technical Approach

3.1 Reader Orientation

LongCat-Image is a complete image generation system built around a 6-billion-parameter diffusion transformer that converts text prompts (in Chinese or English) into high-fidelity images, with specialized capabilities for rendering Chinese characters accurately and performing instruction-based image editing. The system solves a multi-faceted problem — how to achieve photorealistic quality, broad Chinese character coverage, and precise editing consistency without scaling to 20B+ parameters — through an integrated design that spans data curation (filtering billions of samples, generating synthetic text data, producing multi-granularity captions), architectural choices (character-level tokenization for text rendering, hybrid MM-DiT/Single-DiT transformer blocks, 3D multimodal rotary position embeddings), and a five-stage training pipeline (progressive pre-training → mid-training → SFT → DPO → GRPO/MPO) where each stage addresses a specific failure mode of the previous stage.

3.2 Big-Picture Architecture (Diagram in Words)

The LongCat-Image system has seven major components that operate sequentially across training and deployment:

  1. Data Curation Pipeline — ingests raw web-scale images (1.2 billion samples), filters out low-quality, watermarked, AIGC-contaminated, and duplicate content, then enriches remaining images with metadata (categories, style tags, named entities, OCR text, aesthetic scores) and multi-granularity captions, and finally stratifies the corpus into subsets tailored for each training stage.

  2. Text Encoder (Qwen2.5VL-7B) — converts user prompts (text) into continuous vector representations that condition the diffusion model, with a modified character-level tokenization scheme applied specifically to text enclosed in quotation marks (the text to be rendered visually in the generated image) to reduce the model's learning burden for Chinese character glyphs.

  3. VAE (from FLUX.1-dev) — compresses input images from pixel space into a compact latent representation via 8× spatial compression followed by 2×2 token merging, producing latent codes at H/16 × W/16 resolution that serve as the generation target for the diffusion transformer.

  4. Diffusion Transformer (DiT) Backbone — a 30-block transformer with 10 MM-DiT blocks (double-stream attention, where text and image tokens are processed in separate attention streams) followed by 20 Single-DiT blocks (single-stream attention, where text and image tokens are concatenated and processed jointly), using 3D Multimodal Rotary Position Embeddings (M-RoPE) to encode spatial positions and modality identity.

  5. Multi-Stage Training Scheduler — orchestrates the model through five sequential training phases:

    • Pre-training (1.4M steps across three resolutions: 256px → 512px → 512-1024px) on broad data to learn global semantics
    • Mid-training (70K steps) on curated high-fidelity data to refine quality and inject artistic styles
    • SFT (20K steps) on human-curated real+synthetic data to align with aesthetic preferences
    • DPO (4K steps) on human preference pairs to reduce structural failure modes
    • GRPO/MPO (300 steps) with ensemble reward models for fine-grained refinement

  6. Reward Model Ensemble (for RL phases) — a collection of specialized evaluators including an AIGC detector (penalizes synthetic-looking textures), an OCR accuracy model (rewards correct text rendering), aesthetic scoring models (reward visual quality), and distortion detectors (penalize artifacts), whose outputs are combined to compute advantage functions that guide policy updates.

  7. Image Editing Extension — adapts the T2I backbone by adding a reference image conditioning branch (the source image to be edited), initialized from the mid-training checkpoint (not the final aligned model), and trained through three stages (editing pre-training → SFT → DPO) with joint T2I+editing data to prevent catastrophic forgetting of generative knowledge.

Information flows as follows: a prompt enters → the text encoder produces conditioning vectors, with quoted text spans receiving character-level tokenization → these vectors, along with the text's 3D RoPE position embeddings (modality ID=0 for text), are fed into the DiT blocks → the diffusion process starts from random noise latents (modality ID=1 for noise) → at each denoising step, the MM-DiT blocks (first 10 layers) process text and image tokens in separate attention streams, then the Single-DiT blocks (next 20 layers) process them jointly → after all denoising steps, the VAE decoder reconstructs the latent into a pixel-space image → in editing mode, an additional reference image is VAE-encoded (modality ID=2 for reference) and concatenated with the noise latents along the sequence dimension, and the Qwen2.5-VL encoder also receives the reference image alongside the editing instruction.

3.3 Roadmap for the Deep Dive

I will explain the technical approach in six subsections, ordered to build understanding from data foundation to model architecture to training orchestration:

  • First, the data curation pipeline (Section 2.1) — because data quality and stratification are the foundation that enables every subsequent training stage, and the paper's most distinctive technical contributions (AIGC exclusion, multi-granularity captioning, synthetic text data generation, comprehensive aesthetic scoring) are data-centric.

  • Second, the model architecture (Sections 3.1–3.4) — the diffusion transformer backbone, VAE, text encoder with character-level tokenization, and 3D M-RoPE positional embeddings, since these form the fixed computational substrate that the training stages operate on.

  • Third, pre-training (Section 4.1) — the progressive multi-resolution curriculum, dynamic sampling of synthetic text rendering data, and real-time evaluation monitoring, because this is where the model acquires its fundamental text-to-image mapping capabilities.

  • Fourth, mid-training (Section 4.2) — the quality-focused data curation and the rationale for releasing this checkpoint as a "Developer Version," since this stage bridges raw pre-training and aesthetic alignment.

  • Fifth, post-training via SFT and RLHF (Section 4.3) — the SFT data strategy (human-curated real+synthetic, model weight averaging, timestep sampling shift), followed by the three RL algorithms (DPO, GRPO, MPO) with their mathematical formulations, because this is where the model's final capabilities are shaped through preference optimization.

  • Sixth, the image editing adaptation (Sections 6.1–6.3) — data sources, architectural modifications (reference image conditioning branch), and the three-stage training curriculum with its critical design choice of mid-training initialization, because editing shares the T2I backbone but requires fundamentally different data and training strategies.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily a systems and engineering paper whose core idea is that a carefully designed 6B-parameter model, trained on meticulously curated data with stage-appropriate strategies (character-level tokenization for text, AIGC exclusion for photorealism, joint T2I+editing for editing consistency), can match or exceed 20B–80B parameter models by addressing specific failure modes that parameter scaling alone cannot resolve.

Data Curation Pipeline: From Raw Web-Scale Images to Stage-Specific Training Subsets

The data curation pipeline (Section 2.1, Figure 6) transforms a raw corpus of 1.2 billion image-text pairs into distinct training subsets optimized for each training stage, operating in four sequential phases: filtering, metadata extraction, multi-granularity captioning, and stratification.

Phase 1: Filtering. The raw data undergoes five filtering operations applied in sequence. First, deduplication uses a two-tiered strategy: MD5 hashing detects exact duplicates (byte-identical files), while SigLIP-based similarity assessment identifies near-duplicates (visually similar but not byte-identical images). Second, resolution and aspect ratio filtering removes images with a shortest edge below 384 pixels and retains only aspect ratios between 0.25 and 4.0. Third, watermark detection uses a specialized detector (architecture unspecified) to identify and remove samples with visible watermark patterns. Fourth, LAION-Aesthetics filtering evaluates each image with the LAION-Aesthetics predictor and discards images scoring below 4.5 — a threshold "empirically selected to filter out low-quality samples while preserving sufficient diversity for model training." Fifth, AIGC detection uses an internally developed detector to purge AI-generated content from the corpus. This last step is motivated by the empirical observation that AIGC-contaminated data causes models to produce "plastic" or "greasy" textures (Section 2.1.1), and the detector is later reused as a reward model during RL to adversarially push the generator away from synthetic-looking outputs.

Phase 2: Metadata Extraction. Five categories of metadata are extracted from each surviving image. Category assigns the image to one of 16 semantic classes: portrait, sport, activity, plant, animal, food, object, landscape, cityscape, indoor, UI, cartoon, chart, rich-text, poster, and synthetic text. This categorization enables distribution balancing in later stages. Style uses an open-source VLM to produce "plausible style descriptions in the form of phrases, rather than constraining the output to a fixed set of predefined labels" (Section 2.1.2) — this open-vocabulary approach avoids the brittleness of predefined style taxonomies. Named Entity extracts celebrities, fictional characters, biological species, commercial brands, and intellectual properties using VLMs, and this information is fed into the captioning pipeline to improve descriptive accuracy. OCR Text extracts text content from images using an OCR model, which is then integrated into captions with special handling. Comprehensive Aesthetics (Figure 7) decouples evaluation into two orthogonal dimensions — Quality (technical fidelity measured through low-level signal statistics including saturation, contrast, and color richness in RGB and HSV spaces, combined with deep reference-free metrics MUSIQ and Q-Align) and Artistry (photographic merit measured through VLM-based analysis of composition, lighting, shadow, and color tonality, combined with Q-Align-Aesthetics) — using an ensemble of six complementary assessment methodologies.

Phase 3: Multi-Granularity Captioning (MGC). The paper identifies three critical limitations in standard VLM captioning (Section 2.1.3): insufficient integration of world knowledge (named entities described generically), restricted diversity in caption formats (same VLM produces structurally similar captions), and low information density (verbose captions waste token budget). The MGC framework addresses these by organizing captions into four hierarchical levels, each generated by a different model or prompt:

  • Entity Level: Qwen2.5-VL identifies and names the principal visual entities. These are essentially tags — "Hollywood Sign," "young man with dyed white hair," "家居用品推广海报" (home furnishing promotional poster).

  • Phrase Level: The same Qwen2.5-VL call simultaneously produces concise linguistic expressions capturing salient visual attributes — "极简家居陈设，黑色圆形茶几" (minimalist home decor, black round coffee table), "purple sweater, denim jacket, shadow, sunlight."

  • Composition Level: A separate InternVL2.5 model receives the original image, the Entity Level descriptions, and the extracted metadata, and produces an integrative caption capturing overall semantic structure. Example prompt template is shown in Figure 8.

  • Photographic Level: A custom fine-tuned model (the "Photographic Captioner," based on Qwen2.5-VL with LoRA adaptation) produces dense, information-rich descriptions incorporating world knowledge, specific photographic techniques, lighting conditions, and detailed spatial relationships. The LoRA fine-tuning addresses format inconsistency in the base Qwen2.5-VL outputs while preserving the embedded world knowledge (Figure 9 shows a quality comparison where the Photographic Captioner produces more concise and information-dense text than the baseline).

During diffusion model training, these four caption types are sampled with weighted probabilities of [0.05, 0.1, 0.2, 0.65] respectively (Section 2.1.3), heavily favoring the most detailed Photographic Level captions. This distribution "enables the model to accommodate diverse prompt formats while robustly encoding complex world knowledge" — in other words, the model sees enough variety in caption styles to handle arbitrary user prompts at inference time, but the training signal is dominated by the most information-rich captions.

Phase 4: Stratification. The final phase partitions the captioned dataset into subsets optimized for each training stage, using the extracted style tags, semantic categories, and aesthetic scores:

  • Pre-training: Artistic data (illustrations, cartoons, anime) is restricted to approximately 0.5% of the corpus because "exposing the model to a high concentration of artistic data during the early pre-training stage biases the model towards learning simplified visual patterns" and can "compromise the model's ability to generate high-fidelity photorealistic images, effectively causing a 'collapse' in the realistic generation subspace" (Section 2.1.4). The overwhelming majority of pre-training data is photorealistic imagery.

  • Mid-training: Two objectives are pursued. For quality enhancement, a subset of high-resolution images (exceeding 1,024 pixels) is curated from the pre-training corpus, selected for "superior sharpness, balanced composition, and high aesthetic scores." For artistic style injection, previously filtered artistic data is reintroduced with a calibrated schedule that gradually increases the proportion from 0.5% to 2.5%, which "effectively expands the model's stylistic repertoire while preserving its photorealistic foundation."

  • SFT: A hybrid dataset of hundreds of thousands of samples blends human-curated real images (selected by experts evaluating composition, lighting, color tonality, and emotional expression, with balanced categorical distribution) with model-synthesized images that have undergone "rigorous manual filtering to eliminate structural distortions, visual unreality, and aesthetic flaws" (Section 2.1.4). The synthetic data's "strong stylistic consistency" helps the model rapidly converge toward the human preference manifold.

Data Synthesis for Text Rendering (Section 2.2). The Chinese character set's long-tail distribution (roughly 3,000 common characters appearing frequently in natural data, over 5,000 rare characters appearing sparsely) means that conventional training data provides insufficient exposure to rare characters. The paper addresses this by constructing a synthetic text rendering dataset using the SynthDoG tool. The pipeline (Figure 11) renders text from classical literature, Tang poetry, and complex character sequences onto diverse textures (crumpled paper, blackboard, glass) using varied colors, fonts (SimSun, HeiTi, etc.), and layouts (horizontal, vertical), generating over 10 million samples with precise captions specifying the exact text content, color, font, and layout. The captions follow a template like "请在黑板上写一首'千秋岁·数声鶗鴂'，颜色是白色，宋体，横排排版" (Please write the poem '千秋岁·数声鶗鴂' on the blackboard in white, Song font, horizontal layout). Because these synthetic images have monotonous backgrounds designed to "minimize interference and improve the model's focus on learning character glyphs" (Section 1), they are mixed with real images during training and gradually phased out in later stages to prevent overfitting to the synthetic domain.

Model Architecture: Hybrid MM-DiT / Single-DiT Transformer with 3D M-RoPE

The diffusion model architecture (Section 3.1, Figure 12) adopts the transformer backbone from FLUX.1-dev, with specific modifications to the text encoder, tokenization strategy, and positional embeddings.

Diffusion Transformer Backbone. The DiT consists of 30 transformer blocks organized in two sequential groups. The first 10 blocks use MM-DiT (Multi-Modal Diffusion Transformer), a double-stream attention mechanism where text tokens and image (noise latent) tokens are processed in separate attention streams within each block — the queries, keys, and values for text and image modalities are computed independently, allowing the model to learn modality-specific transformations before mixing information. The remaining 20 blocks use Single-DiT, a single-stream attention mechanism where text and image tokens are concatenated along the sequence dimension and processed jointly through standard self-attention. The ratio of double-stream to single-stream blocks is maintained at approximately 1:2 to "ensure parameter balance" (Section 3.1).

VAE. The Variational Autoencoder is taken directly from FLUX.1-dev without modification. Input images undergo 8× spatial compression in the encoder, producing latent representations at 1/8 the original spatial resolution. These latents are further processed through 2×2 token merging (patchification), yielding a final sequence length of (H×W)/(16×16) tokens that enter the DiT module. For a 1024×1024 image, this produces 64×64 = 4096 latent tokens. The paper notes that empirical evaluations demonstrate this VAE's "superior reconstruction fidelity in challenging scenarios, such as fine typography and intricate textures" (Section 3.1).

Text Encoder and Character-Level Tokenization (Section 3.2). The text encoder is Qwen2.5VL-7B, a 7-billion-parameter vision-language model that processes user prompts (text only, during T2I generation) into continuous vector representations. The paper departs from prior models that used CLIP and T5 encoders, noting that recent work has shifted toward LLMs or MLLMs "to enhance multilingual compatibility, particularly for Chinese" (Section 3.2). A key architectural decision is the abandonment of "the conventional injection of text embeddings into timestep embeddings for adaLN modulation," since empirical evidence suggested negligible performance gains from this operation.

The critical innovation for text rendering is character-level tokenization for quoted text spans. During tokenization, text enclosed in quotation marks in the prompt (the content to be rendered visually in the generated image) is processed at character granularity rather than through the standard subword tokenizer. Each Chinese character becomes a separate token. The paper states this "mitigates generation complexity without incurring the computational costs and memory footprint of specialized encoders (e.g., GlyphByT5)" and "not only improves data efficiency but also accelerates convergence for text rendering tasks" (Section 3.2).

The mechanism behind this improvement is worth unpacking. Standard subword tokenizers like those in LLMs may split a rare Chinese character into multiple subword units, meaning the diffusion model must learn to render a visual glyph from a fragmented conditioning signal spread across multiple tokens' embeddings. For common characters, this fragmentation is manageable because the model sees many examples. For rare characters (5,000+ in the Chinese lexicon), the combinatorial explosion of subword fragments makes learning essentially impossible without massive data scale. Character-level tokenization eliminates this fragmentation: each character corresponds to exactly one token embedding, and the model learns a direct mapping from that embedding to the visual glyph. This reduces the learning problem from "compose a glyph from fragmented subword embeddings" to "memorize a glyph per character token," which is feasible with modest synthetic data.

3D Multimodal Rotary Position Embeddings (M-RoPE) (Section 3.3). Positional embeddings encode the spatial location of each token — critical for diffusion models that generate images at variable resolutions and aspect ratios. The paper adopts vanilla 3D M-RoPE without modification, citing empirical observations that "the model possesses intrinsic adaptability to varying positional strides across different resolutions, rendering these explicit geometric constraints unnecessary" (Section 3.3). This contrasts with prior work (PixArt-α, Hunyuan-DiT, Seedream 2.0) that employed "intricate heuristics — such as coordinate centering, frequency scaling, or interpolation — to align spatial distributions."

The 3D M-RoPE uses three dimensions with distinct purposes. The first dimension encodes modality identity: for text-to-image tasks, distinct values distinguish noise latent tokens from text tokens. For image editing tasks, this dimension further distinguishes reference image latents from the two aforementioned types. The remaining two dimensions encode 2D spatial coordinates: for image tokens, they correspond to the (x, y) position within the latent grid. For text tokens, both coordinates are set to the same value, equivalent to 1D-RoPE behavior. This design enables "flexible image generation across arbitrary aspect ratios" and "seamless interaction with other modalities" (Section 3.3).

Prompt Engineering (Section 3.4). To bridge the gap between training captions (dense, detailed, following the Photographic Level format) and user prompts (concise, often ambiguous), the paper provides a built-in prompt refinement solution that repurposes the existing Qwen2.5-VL text encoder as a prompt rewriter. This eliminates dependencies on external APIs or LLMs while "ensuring an out-of-the-box capability for generating high-fidelity images" (Section 3.4). The exact rewriting mechanism — whether it's fine-tuned for this purpose, uses a specific system prompt, or employs in-context learning — is not detailed.

Pre-training: Progressive Multi-Resolution Curriculum with Dynamic Synthetic Data Sampling

Pre-training (Section 4.1) is where the model acquires its fundamental text-to-image mapping capability. The training operates over 1.4 million steps across three resolution stages, with a dynamic strategy for incorporating synthetic text rendering data.

Progressive Mixed-Resolution Training. The pre-training curriculum spans three sequential phases, each with distinct resolutions and learning rates (Table 1):

  • Stage 1 (256px): 900K steps, learning rate 1e-4, global batch size 4608. This stage focuses on efficient acquisition of global semantic knowledge — the model learns high-level concepts (what a "dog" or "sunset" looks like) at low resolution where training is fast.

  • Stage 2 (512px): 300K steps, learning rate 5e-5, global batch size 4608. This explicit intermediate stage is retained to "avoid the computational instability of transitioning directly from 256px to the final phase" (Section 4.1). The model refines its understanding at medium resolution before tackling high-resolution details.

  • Stage 3 (512–1024px, dynamic): 200K steps, learning rate 2e-5, global batch size 3072. This stage covers a continuous resolution range, with bucket sampling to accommodate variable aspect ratios. The model learns high-frequency detail refinement at resolutions up to 1024px.

All stages use the AdamW optimizer with β1 = 0.9, β2 = 0.95, weight decay 0.01, gradient clipping at 1.0, and constant learning rate schedules (no warmup, no decay within each stage). The progressive reduction in learning rate across stages (1e-4 → 5e-5 → 2e-5) reflects a standard practice of using higher learning rates for coarse feature learning and lower rates for fine detail refinement.

Dynamic Sampling of Synthetic Text Rendering Data. The synthetic text dataset (10M+ samples from SynthDoG) presents a tradeoff: it significantly enhances text rendering accuracy by providing concentrated exposure to character glyphs, but "inevitably compromises the overall visual harmony" (Section 4.1) because the simplistic backgrounds (paper, glass, blackboard) don't teach the model how to integrate text naturally into complex scenes. The solution is a two-pronged dynamic strategy:

  1. Character-wise accuracy monitoring: During training, the system tracks the error rate for each Chinese character in real-time using OCR evaluation on generated samples. Characters with high error rates (the model consistently fails to render them correctly) have their sampling probability increased in the synthetic data stream, ensuring the model receives more training on its weakest characters.

  2. Progressive phase-out: For characters that achieve low error rates (well-learned), the synthetic sampling probability is gradually reduced in favor of real images containing those characters in natural contexts. In the final stage of pre-training, synthetic data is "completely phased out" to prevent the model from overfitting to the simplistic synthetic domain and to ensure it learns to render text that integrates naturally with complex backgrounds.

This dynamic curriculum directly addresses the long-tail character distribution: the model gets concentrated synthetic exposure to rare characters (which appear too infrequently in natural data to learn from) during early and mid pre-training, then transitions to real-world text-in-image data to learn natural integration once the basic glyph forms are established.

Real-time Evaluation Protocol. A comprehensive monitoring system tracks validation loss, image-text alignment, aesthetic scores, and OCR-based text rendering accuracy throughout pre-training. The paper notes that while these metrics "serve as pivotal indicators during the pre-training phase, their discriminative power and sensitivity notably diminish during the mid-training and post-training stages as performance saturates" (Section 4.1). This is an important practical observation: evaluation metrics that effectively distinguish model quality during early training become saturated and uninformative once the model reaches a certain quality threshold, necessitating different evaluation strategies (human evaluation, benchmark suites) in later stages.

Mid-training: Quality-Constrained Fine-Tuning on Curated High-Fidelity Data

Mid-training (Section 4.2) serves as a bridge between the broad but noisy pre-training distribution and the narrow aesthetic manifolds targeted by post-training alignment. The paper's framing is that pre-training produces "robust global semantic priors and text-to-image mapping capabilities" but the "resulting visual outputs often lack high-fidelity textures and aesthetic coherence due to the inherent noise in large-scale pre-training data" (Section 4.2). Mid-training constrains the learned manifold toward higher quality without yet locking the model into a specific aesthetic style.

Data Curation. The mid-training dataset is curated through a "rigorous data curation protocol that is significantly more stringent than that of the pre-training phase" (Section 4.2). The pipeline integrates a "hierarchical assessment system comprising advanced aesthetic scoring models, image quality estimators, and domain-specific classifiers, culminating in a human-in-the-loop verification process." This yields "a high-fidelity corpus of millions of samples, ensuring balanced representation across diverse domains" (Section 4.2).

The key metrics are sharpness, balanced composition, and high aesthetic scores — essentially, the mid-training data represents the subset of pre-training data that passes stricter quality thresholds. The artistic style injection is achieved by reintroducing previously filtered artistic data (illustrations, cartoons, anime) at a gradually increasing proportion from 0.5% to 2.5%, which expands the model's stylistic range while keeping the overwhelming majority of training data photorealistic.

Training Configuration. Mid-training runs for 70K steps with learning rate 1e-5, global batch size 3072, and a constant learning rate schedule (Table 1). The optimizer configuration (AdamW, β1=0.9, β2=0.95, weight decay 0.01, gradient clip 1.0) is identical to pre-training.

Developer Version Release. The checkpoint from this stage is designated as the "Developer Version" and released to the community. The rationale (Section 4.2) is that this model "retains high plasticity and adaptability, avoiding the potential mode collapse or rigidity often introduced by aggressive alignment" during subsequent SFT and RL stages. In practice, this means the mid-training checkpoint is the appropriate starting point for downstream fine-tuning on specific domains or styles, because it has broad generative capabilities without being locked into the specific aesthetic preferences encoded during SFT. This is the same checkpoint used to initialize the image editing model (Section 6.3.1), where the unconstrained parameter space offers "superior plasticity compared to post-trained models" for learning diverse editing tasks.

Post-training via SFT: Aesthetic Alignment Through Curated Data, Model Merging, and Timestep Resampling

The SFT phase (Section 4.3.1) shifts the objective from broad quality enhancement (mid-training) to targeted aesthetic alignment with human preferences. It operates for 20K steps with learning rate 1e-5, global batch size 128 (dramatically reduced from 3072 in mid-training, reflecting the much smaller curated dataset), a cosine learning rate schedule (rather than constant), and 1000 warmup steps. Three key techniques distinguish this SFT phase from standard fine-tuning.

High-Fidelity Data Curation. The SFT dataset comprises "hundreds of thousands of samples" blending two sources. Real-world imagery is manually curated by human experts who evaluate "dimensions such as composition, lighting, color tonality, and emotional expression, while ensuring a balanced categorical distribution" (Section 4.3.1). High-quality synthetic images are generated by the model itself (or other models) and undergo "rigorous manual filtering to eliminate structural distortions, visual unreality, and aesthetic flaws" (Section 2.1.4). The synthetic data's value is its "strong stylistic consistency," which accelerates convergence toward the human preference manifold — but the paper explicitly notes that a "strict expert verification protocol" is enforced to prevent "the degradation of realism often associated with synthetic artifacts" (Section 4.3.1). Only data "possessing superior textural quality and aesthetic value" is retained.

Model Weight Averaging. Rather than training a single SFT model on all data, the paper fine-tunes "multiple candidate models where each is specialized in distinct visual dimensions, including illumination, portraiture, and artistic style" (Section 4.3.1). These specialized models are trained on different subsets of the SFT data, each emphasizing different aesthetic attributes. The final SFT model is produced by weight averaging (merging the parameters of all specialized models, presumably through simple arithmetic mean of corresponding weights, though the exact merging formula is not specified). This strategy "effectively balances performance across multiple attributes" and "significantly enhances overall robustness and stability, effectively mitigating the specific biases or deficits inherent in individual single-domain models" (Section 4.3.1). The intuition is that each specialized model overfits to its narrow domain (e.g., portrait lighting) in slightly different ways, and averaging cancels out these idiosyncratic biases while preserving the shared improvements.

Optimization of Timestep Sampling. The pre-training phase used Logit-Normal sampling for timesteps during diffusion training, which concentrates training on intermediate timesteps (where the model learns global structure). SFT switches to Uniform Sampling across all timesteps. The rationale (Section 4.3.1): "SFT focuses on refining high-frequency details that typically emerge during the later stages of the diffusion process" (i.e., low noise levels, where the model adds final textures and fine details). Uniform sampling "ensures balanced exposure across all timesteps, specifically increasing the training weight of high-frequency denoising steps to maximize the model's capacity for learning intricate textures and fine details." In other words, Logit-Normal sampling under-trains the low-noise regime where photorealistic details are finalized; uniform sampling corrects this imbalance for the SFT phase where detail quality is the primary target.

Post-training via RLHF: DPO, GRPO, and MPO with Ensemble Reward Models

The RLHF phase (Section 4.3.2) applies three distinct reinforcement learning algorithms — DPO, GRPO, and MPO — sequentially, each serving a different purpose in the alignment pipeline. All are initialized from the SFT checkpoint and use ensemble reward models that evaluate images on multiple quality dimensions simultaneously.

Reward Model Ensemble. The paper develops "fine-grained reward models (RMs), including distortion detection, AIGC detection, human preference assessment, and OCR accuracy, to comprehensively evaluate the model's detailed capabilities" (Section 4.3.2). The specific models are:

  • Distortion detection: Identifies structural artifacts, anatomical errors, and unnatural proportions.
  • AIGC detection: The same internally developed detector used in data filtering, now repurposed as an adversarial reward signal — images that look AI-generated (plastic/synthetic textures) receive lower rewards, pushing the model toward photorealism.
  • Human preference assessment: Evaluates overall aesthetic quality, composition, lighting, and color harmony, likely trained on human annotation data.
  • OCR accuracy: Measures whether text rendered in generated images is correct and legible, providing a direct signal for text rendering quality.

These RMs produce scalar scores that are combined (the exact combination method — weighted sum, minimum, or learned aggregation — is not specified) to compute the reward for each generated image.

Direct Preference Optimization (DPO). DPO is applied first, at relatively large scale, to address "common structural deficiencies in the model" (Section 4.3.2, DPO subsection). It operates for 4K steps with learning rate 1e-5, global batch size 64, and cosine learning rate schedule with 1000 warmup steps (Table 1).

Data Construction for DPO. The preference dataset is built in two steps. First, a PromptSet is constructed from "real user queries of public datasets, refined through clustering and data-cleaning techniques to ensure representativeness and diversity" (Section 4.3.2, DPO subsection). Second, for each prompt, 6 candidate images are generated using diverse random initialization seeds. A human annotation team assigns subjective quality scores on a 1–5 scale. Neutral samples (score = 3) are discarded to ensure high-confidence preference signals; high-quality samples (scores 4–5) are treated as positive (winner, $x^w_0$), and low-quality samples (scores 1–2) are treated as negative (loser, $x^l_0$), forming win-lose pairs for training.

DPO Objective. The DPO loss for flow-matching models is:

L(θ)=−E(x0w,x0l)∼D,t∼U(0,T),xtw∼q(xtw∣x0w),xtl∼q(xtl∣x0l)log⁡σ(−βTω(λt)[∥vw−vθ(xtw,t)∥22−∥vw−vref(xtw,t)∥22−(∥vl−vθ(xtl,t)∥22−∥vl−vref(xtl,t)∥22)])\mathcal{L}(\theta) = -\mathbb{E}_{(x^w_0, x^l_0) \sim \mathcal{D}, t \sim \mathcal{U}(0,T), x^w_t \sim q(x^w_t|x^w_0), x^l_t \sim q(x^l_t|x^l_0)} \log \sigma \left( -\beta T \omega(\lambda_t) \left[ \|v^w - v_\theta(x^w_t, t)\|^2_2 - \|v^w - v_{\text{ref}}(x^w_t, t)\|^2_2 - \left( \|v^l - v_\theta(x^l_t, t)\|^2_2 - \|v^l - v_{\text{ref}}(x^l_t, t)\|^2_2 \right) \right] \right)

where $x^w_0, x^l_0$ are the winning and losing images from the preference pair, $\mathcal{D}$ is the preference dataset, $t \sim \mathcal{U}(0,T)$ is a uniformly sampled timestep, $q(x_t|x_0)$ is the forward diffusion process that adds noise to the clean image, $\sigma(\cdot)$ is the sigmoid function, $\beta$ is a temperature parameter controlling the strength of preference optimization, $\omega(\lambda_t)$ is a timestep-dependent weighting function (related to the signal-to-noise ratio $\lambda_t$), $v^w, v^l$ are the target velocity fields for the winning and losing samples respectively, $v_\theta$ is the model's predicted velocity field, and $v_{\text{ref}}$ is the reference model's (SFT checkpoint) velocity field.

What it computes: For each preference pair, the loss compares how much better the current model $\theta$ is than the reference model at predicting the winning sample's velocity field versus the losing sample's. Specifically, the term $\|v^w - v_\theta(x^w_t, t)\|^2_2 - \|v^w - v_{\text{ref}}(x^w_t, t)\|^2_2$ measures whether the current model has improved (reduced error) relative to the reference on the winning sample — a negative value means improvement. Similarly, the losing sample term $\|v^l - v_\theta(x^l_t, t)\|^2_2 - \|v^l - v_{\text{ref}}(x^l_t, t)\|^2_2$ measures improvement on the losing sample. The difference of these differences (winning improvement minus losing improvement) is scaled by $-\beta T\omega(\lambda_t)$, passed through the sigmoid, and the log is taken. The loss is minimized when the model improves more on winning samples than on losing samples — i.e., when it shifts its distribution toward higher-quality outputs.

Why this form: This is the standard DPO objective adapted for flow-matching models, where the implicit reward is the negative squared error between predicted and target velocity fields (rather than the log-probability ratio used in LLM DPO). The reference model term provides an implicit KL constraint — the model is penalized for deviating too far from the SFT checkpoint's predictions, which prevents reward hacking where the model would learn to produce outputs that maximize the reward model's score but are visually degenerate. The paper notes that during training, they "further explore strategies including multi-round DPO iterations, gradient norm-based dirty data skipping, and KL constraints" (Section 4.3.2, DPO subsection), with the ultimate finding that "DPO significantly reduces the model's bad case rate and enhances the robustness of image generation."

Group Relative Policy Optimization (GRPO). GRPO is applied after DPO for "further fine-grained training" (Section 4.3.2, GRPO subsection). It operates for 300 steps with learning rate 5 × 10^{-6} (lower than DPO's 1e-5), global batch size 64, and no warmup (Table 1: warmup steps = 0). Unlike DPO (offline, uses pre-collected preference pairs), GRPO performs on-policy sampling — the model generates images during training, evaluates them with reward models, and updates based on those evaluations.

GRPO Sampling Process. Following the Dance-GRPO framework, for each text prompt with hidden state $h$, the flow model generates a group of $G$ images $\{x^i_0\}^G_{i=1}$ along with their corresponding denoising trajectories $\{x^i_T, x^i_{T-1}, ..., x^i_0\}^G_{i=1}$. The generation uses a reformulated stochastic differential equation (SDE) to enable exploration (rather than the deterministic ODE used at inference time):

dxt=(vt+σt22t(xt+(1−t)vt))dt+σtdwdx_t = \left( v_t + \frac{\sigma^2_t}{2t}(x_t + (1-t)v_t) \right) dt + \sigma_t dw

where $v_t$ is the deterministic flow-matching velocity, $\sigma_t$ is a time-dependent diffusion coefficient enabling stochastic exploration, and $dw$ is standard Brownian motion. Discretized via Euler-Maruyama:

xt+Δt=xt+(vθ(xt,t,h)+σt22t(xt+(1−t)vθ(xt,t,h)))Δt+σtΔtϵx_{t+\Delta t} = x_t + \left( v_\theta(x_t, t, h) + \frac{\sigma^2_t}{2t}(x_t + (1-t)v_\theta(x_t, t, h)) \right) \Delta t + \sigma_t \sqrt{\Delta t} \epsilon

where $\epsilon \sim \mathcal{N}(0, I)$ is Gaussian noise. The SDE formulation is critical: the deterministic ODE would always produce the same image for a given noise seed, which provides no diversity for comparing within-group quality. The stochastic term $\sigma_t dw$ injects controlled randomness, producing slightly different images that can be compared and ranked.

GRPO Advantage and Objective. Within each group of $G$ images, the advantage function for the $i$-th image is:

Ai=R(x0i,h)−mean({R(x0j,h)}j=1G)std({R(x0j,h)}j=1G)A^i = \frac{R(x^i_0, h) - \text{mean}(\{R(x^j_0, h)\}^G_{j=1})}{\text{std}(\{R(x^j_0, h)\}^G_{j=1})}

where $R(x^i_0, h)$ is the reward (ensemble RM score) for image $i$ given prompt encoding $h$, $\text{mean}(\cdot)$ is the average reward within the group, and $\text{std}(\cdot)$ is the standard deviation within the group.

What it computes: The advantage $A^i$ is a normalized score measuring how much better (or worse) image $i$ is compared to the average image in its group, in units of within-group standard deviation. An advantage of +1.0 means the image is one standard deviation better than the group mean; -0.5 means half a standard deviation worse. This group-relative normalization is critical because absolute reward scales can vary dramatically across prompts (some prompts are inherently harder, yielding lower absolute rewards even for good images). By normalizing within each prompt's group, GRPO creates a consistent advantage scale across diverse prompts.

Why this form: The group-relative normalization acts as an automatic baseline that adapts to prompt difficulty. Without it, the policy gradient would be dominated by easy prompts (where rewards are consistently high) and noise-dominated for hard prompts (where rewards are low and variance is high). The z-score normalization equalizes the effective learning signal across prompts.

The GRPO training objective is:

LGRPO(θ)=Eh∼D,{xTi,...,x0i}i=1G∼πθ[1G∑i=1G1T∑t=0T−1min⁡(rti(θ)Ai,clip(rti(θ),1−ϵ,1+ϵ)Ai)]\mathcal{L}_{\text{GRPO}}(\theta) = \mathbb{E}_{h \sim \mathcal{D}, \{x^i_T, ..., x^i_0\}^G_{i=1} \sim \pi_\theta} \left[ \frac{1}{G} \sum_{i=1}^G \frac{1}{T} \sum_{t=0}^{T-1} \min\left( r^i_t(\theta) A^i, \text{clip}(r^i_t(\theta), 1-\epsilon, 1+\epsilon) A^i \right) \right]

where $r^i_t(\theta) = \frac{p_\theta(x^i_{t-1}|x^i_t, h)}{p_{\theta_{\text{old}}}(x^i_{t-1}|x^i_t, h)}$ is the probability ratio between the current policy and the old policy (from the previous iteration) for the transition from timestep $t$ to $t-1$, and $\epsilon$ is the clipping threshold (typically 0.1–0.2, following standard PPO practice, though the exact value is not specified).

What it computes: For each image in the group, at each denoising timestep, the objective computes the probability ratio $r^i_t$ — how much more (or less) likely the current model is to make the same denoising step compared to the old model — and multiplies it by the advantage $A^i$. The min with clipping prevents the policy from changing too drastically: if $r^i_t$ exceeds $1+\epsilon$ (model became much more likely) for a positive advantage, the clipped value $1+\epsilon$ is used instead, preventing the model from overfitting to a single high-reward trajectory. If $r^i_t$ falls below $1-\epsilon$ (model became much less likely) for a negative advantage, the clipped value $1-\epsilon$ is used, preventing the model from catastrophically forgetting reasonable trajectories.

Why this form: This is the standard PPO clipped objective, adapted for flow-matching by replacing the log-probability ratio with the flow-matching probability ratio $r^i_t$. The clipping mechanism is the key innovation of PPO over earlier policy gradient methods — it provides a trust region that prevents destructively large policy updates while still allowing the model to improve. The per-timestep application (summing over $t$ from 0 to $T-1$) means the policy is optimized across the entire denoising trajectory, not just the final output.

Implementation details for GRPO (Section 4.3.2, Training Strategy subsection): GRPO uses a 12-step SDE sampler (i.e., $T=12$) with Euler-Maruyama discretization. The diffusion coefficient $g_t$ (equivalent to $\sigma_t$ in the SDE formulation) is linearly annealed from 0.1 to 0 during training — meaning exploration is highest at the start of the denoising trajectory and gradually reduced, converging to the deterministic ODE at the final timestep. This annealing schedule balances exploration (needed to discover diverse candidates for within-group comparison) with exploitation (needed to produce high-quality images that achieve good rewards). These experiments are initialized from the DPO checkpoint.

Monolithic Policy Optimization (MPO). MPO is presented as an improvement over GRPO that "fundamentally improves upon GRPO by eliminating the group-relative paradigm and its associated synchronization bottlenecks, achieving superior training efficiency and stability" (Section 4.3.2). The key insight is that GRPO requires generating a group of $G$ images per prompt (creating a synchronization barrier — all $G$ must complete before the advantage can be computed and the update applied), whereas MPO generates a single trajectory per prompt and performs one gradient update immediately.

MPO Sampling and SDE. MPO uses the same SDE formulation as GRPO for generation:

dz=vθ(zt,c,t)dt+g(t)dwdz = v_\theta(z_t, c, t) dt + g(t) dw

Discretized via Euler-Maruyama:

zt+Δt=zt+vθ(zt,c,t)Δt+gtΔtϵt,ϵt∼N(0,I)z_{t+\Delta t} = z_t + v_\theta(z_t, c, t) \Delta t + g_t \sqrt{\Delta t} \epsilon_t, \quad \epsilon_t \sim \mathcal{N}(0, I)

The critical difference is that only one trajectory is generated per prompt (no group), so there's no within-group comparison to normalize the advantage. MPO must therefore maintain its own baseline for computing advantages.

MPO Advantage Estimation: Gaussian Value Tracker with KL-Adaptive Forgetting. MPO maintains a per-prompt value estimate $V(c) = \mathcal{N}(\mu_c, \sigma^2_c)$ — a Gaussian distribution representing the expected reward for prompt $c$ and the uncertainty about that expectation. The mean $\mu_c$ serves as the baseline (subtracted from the observed reward to compute advantage), and the variance $\sigma^2_c$ quantifies epistemic uncertainty (how confident we are in the baseline). Updates follow Kalman filter principles:

Kt=σc,t−12σc,t−12+σobs2K_t = \frac{\sigma^2_{c,t-1}}{\sigma^2_{c,t-1} + \sigma^2_{\text{obs}}}

μc,t←μc,t−1+Kt(r−μc,t−1)\mu_{c,t} \leftarrow \mu_{c,t-1} + K_t (r - \mu_{c,t-1})

σc,t2←(1−Kt)σc,t−12+Qt\sigma^2_{c,t} \leftarrow (1 - K_t) \sigma^2_{c,t-1} + Q_t

where $K_t$ is the Kalman gain (determining how much to trust the new observation vs. the prior estimate), $\sigma^2_{\text{obs}}$ is the observation noise variance, $r$ is the observed reward for the current trajectory, and $Q_t = \alpha \cdot D_{\text{KL}}(\pi_{\theta'} \| \pi_\theta)$ is an adaptive process noise term that scales with the KL divergence between the current policy $\pi_{\theta'}$ and the reference policy $\pi_\theta$ (i.e., when the policy has changed significantly since the last value estimate update, we increase uncertainty to reflect that the old baseline may be stale).

What it computes: The Kalman update produces a running estimate of the expected reward per prompt. When a new reward $r$ is observed, the mean $\mu_c$ moves toward $r$ by an amount proportional to the Kalman gain $K_t$. If uncertainty $\sigma^2_c$ is high (we're not confident in our baseline), $K_t$ is close to 1 and the baseline rapidly adapts. If uncertainty is low, $K_t$ is close to 0 and the baseline changes slowly. The process noise $Q_t$ injects additional uncertainty when the policy shifts, preventing the baseline from becoming overconfident and stale.

Why this form: Standard baselines in policy gradient methods (e.g., running average of rewards) have no notion of uncertainty and no mechanism to detect when the baseline has become stale due to policy drift. The Kalman filter approach provides both: $\mu_c$ tracks the expected reward, and $\sigma^2_c$ tracks when we should trust vs. discount it. The KL-adaptive forgetting term $Q_t$ is particularly important because MPO updates the policy after every single trajectory, meaning the policy can shift rapidly, and a baseline that doesn't account for this shift would provide misleading advantages.

Global Advantage Normalization. The raw advantage $A = r - \mu_c$ is further normalized using exponential moving averages (EMA) of the mean and variance of advantages across all prompts:

A~=A−μAσA2+ϵ\tilde{A} = \frac{A - \mu_A}{\sqrt{\sigma^2_A + \epsilon}}

where $\mu_A$ and $\sigma^2_A$ are EMA estimates of the global advantage mean and variance, and $\epsilon$ is a small constant for numerical stability. This ensures that advantages are on a consistent scale regardless of prompt difficulty or reward scale.

Uncertainty-Powered Curriculum. MPO samples prompts with probability proportional to $p(c) \propto \sigma_c + \eta / \sqrt{n_c + 1}$, where $\sigma_c$ is the uncertainty for prompt $c$, $n_c$ is the number of times prompt $c$ has been sampled, and $\eta$ is a balance coefficient (set to 1.0). This prioritizes high-uncertainty prompts (where the model's performance is poorly estimated, suggesting room for improvement) while ensuring some exploration of all prompts through the $\eta/\sqrt{n_c+1}$ term.

MPO Policy Update. The policy is updated using advantage-weighted regression:

LMPO(θ)=Et,zt∼τ[stop_grad(wc⋅A~)⋅∥vθ(zt,c,t)−u(zt,z0)∥2]\mathcal{L}_{\text{MPO}}(\theta) = \mathbb{E}_{t, z_t \sim \tau} \left[ \text{stop\_grad}(w_c \cdot \tilde{A}) \cdot \| v_\theta(z_t, c, t) - u(z_t, z_0) \|^2 \right]

where $u(z_t, z_0)$ is the target flow-matching vector field (the direction from the noised latent $z_t$ to the clean latent $z_0$), $v_\theta(z_t, c, t)$ is the model's predicted velocity, and $w_c = 1 + \gamma \cdot |r - \mu_c| / (\sigma_c + \epsilon)$ is a surprise reweighting factor (with $\gamma = 0.5$) that scales the advantage by how surprising the observed reward is relative to the expected distribution.

What it computes: The loss is a weighted mean squared error between the model's predicted velocity field and the target velocity field, where the weight is the normalized advantage $\tilde{A}$ multiplied by the surprise factor $w_c$. The stop_grad operator means the weight is treated as a constant during backpropagation (the gradient only flows through the MSE term). When $\tilde{A}$ is positive (image was better than expected), the weight is positive and the model is encouraged to reproduce the velocity field that led to this good image. When $\tilde{A}$ is negative (image was worse than expected), the model is discouraged from reproducing that velocity field.

Why this form: This is a simple but effective policy gradient variant: instead of the clipped probability ratio used in PPO/GRPO, MPO directly weights the regression loss by the advantage. The surprise reweighting $w_c$ amplifies the learning signal for unexpectedly good or bad outcomes (large $|r - \mu_c|$) and dampens it for outcomes close to expectation — this focuses optimization on the trajectories that provide the most information about what the model is doing right or wrong.

MPO-specific hyperparameters (Section 4.3.2, Training Strategy subsection): EMA decay for advantage normalization $\lambda = 0.99$, curriculum balance coefficient $\eta = 1.0$, adaptive scaling factor $\alpha = 1.0$, surprise reweighting factor $\gamma = 0.5$. These experiments are initialized from the same DPO checkpoint as GRPO, use the same optimizer configuration, and employ a 12-step SDE sampler with $g_t$ linearly annealed from 0.1 to 0.

DPO for Editing (Section 6.3.3). The editing model also undergoes DPO training, using an analogous objective but conditioned on the source image and editing prompt. The DPO loss for editing is:

L(θ)=−E(Isrcw,Pw,Isrcl,Pl)∼D,t∼U(0,T),xtw∼q(xtw∣Isrcw,Pw),xtl∼q(xtl∣Isrcl,Pl)log⁡σ(−βTω(λt)[∥vw−vθ(xtw,Isrcw,Pw,t)∥22−∥vw−vref(xtw,Isrcw,Pw,t)∥22−(∥vl−vθ(xtl,Isrcl,Pl,t)∥22−∥vl−vref(xtl,Isrcl,Pl,t)∥22)])\mathcal{L}(\theta) = -\mathbb{E}_{(I^w_{\text{src}}, P^w, I^l_{\text{src}}, P^l) \sim \mathcal{D}, t \sim \mathcal{U}(0,T), x^w_t \sim q(x^w_t | I^w_{\text{src}}, P^w), x^l_t \sim q(x^l_t | I^l_{\text{src}}, P^l)} \log \sigma \left( -\beta T \omega(\lambda_t) \left[ \|v^w - v_\theta(x^w_t, I^w_{\text{src}}, P^w, t)\|^2_2 - \|v^w - v_{\text{ref}}(x^w_t, I^w_{\text{src}}, P^w, t)\|^2_2 - \left( \|v^l - v_\theta(x^l_t, I^l_{\text{src}}, P^l, t)\|^2_2 - \|v^l - v_{\text{ref}}(x^l_t, I^l_{\text{src}}, P^l, t)\|^2_2 \right) \right] \right)

where $I^w_{\text{src}}, I^l_{\text{src}}$ are the source images for the winning and losing editing pairs, and $P^w, P^l$ are the corresponding editing prompts. The structure is identical to the T2I DPO objective but Conditions the velocity prediction on both the source image and the editing instruction. The editing DPO data construction follows the same two-step procedure: prompt curation from MLLM outputs and user queries, followed by human annotation of 5 candidates per prompt to identify winners and losers. Advanced strategies — "gradient-based outlier rejection (to skip noisy data) and KL divergence constraints" (Section 6.3.3) — are applied to stabilize training.

Image Editing Architecture and Training: Reference Image Conditioning with Mid-Training Initialization

The image editing capability (Section 6) is built by extending the T2I architecture with a reference image conditioning branch, trained through a separate three-stage pipeline. The critical design decisions are the architectural modifications, the choice to initialize from mid-training rather than post-training checkpoints, and the joint training strategy with T2I data.

Architectural Modifications (Section 6.2, Figure 20). The editing model adds an image conditioning branch that processes the source image (the image to be edited) in parallel with the noise latents. Specifically:

  • VAE Encoding: The source image is encoded through the same VAE as the noise latents, producing a second set of latent tokens at the same spatial resolution.

  • 3D RoPE Modality Differentiation: The first dimension of the 3D M-RoPE embeddings is extended to include a new modality ID: 0 for text tokens, 1 for noise latent tokens, and 2 for reference image tokens. The remaining two dimensions encode the same spatial coordinates as the noise latents (preserving spatial alignment — a reference token at position (x,y) has the same spatial embedding as the noise latent at (x,y)).

  • Token Concatenation: The reference image tokens are concatenated with the noise latent tokens along the sequence dimension before being fed into the DiT blocks. This means the DiT sees a combined sequence of [reference_tokens; noise_tokens] for the visual stream, plus the text tokens from the encoder.

  • Multimodal Encoder Integration: The Qwen2.5-VL text encoder receives both the source image and the editing instruction (unlike T2I mode where it only receives text). A distinct system prompt differentiates editing tasks from standard T2I generation during feature extraction.

Training Data Curation for Editing (Section 6.1). The editing training set is constructed from five diverse sources:

  1. Open-Source Datasets: OmniEdit, OmniGen2, and NHREdit are leveraged as high-quality base sources. The paper implements a "rigorous data cleaning pipeline" on these datasets and rewrites the original instructions to improve diversity.

  2. Synthesized Data: Specialized expert models are used to generate training pairs for specific tasks: object manipulation, style transfer, background alteration, and reference-based generation. For each task, MLLMs craft editing instructions, expert models produce target images, and traditional image processing algorithms handle low-level adjustments (filter transformations, lighting changes). Human-in-the-loop verification ensures semantic alignment and visual fidelity.

  3. Video Frames: To capture complex structural changes (human pose, perspective) that synthetic methods struggle with, keyframe pairs are extracted from video sequences. Multimodal models identify target objects, optical flow quantifies changes between frames, and pairs exhibiting "significant yet coherent variations" are automatically annotated with editing instructions. A subset undergoes manual verification.

  4. Interleaved Corpus: Web-scale image-text sequences with "inherent semantic correlations" are mined for implicit editing signals. Raw pairs undergo "rigorous filtering and multimodal-assisted instruction rewriting." The paper acknowledges that this is "an extremely resource-intensive endeavor" and that "the scale of data we have curated to date remains limited" (Section 6.1.4), suggesting this is a work-in-progress direction.

  5. Instruction Rewriting: GPT-4o is used to enhance instruction diversity through a one-to-many strategy — each editing pair is associated with "multiple rewritten variants, including natural language paraphrases and compound commands" (Section 6.1.5). This "aligns training data with diverse inference scenarios."

Figure 19 shows the task distribution, with the largest categories being: Subject Replacement (9.1%), Subject Addition (7.8%), Subject Removal (9.3%), Text Modification (11.3%), View Change (9.2%), and Hybrid Change (9.8%).

Editing Training Pipeline (Section 6.3). The editing model is trained in three stages:

Stage 1: Pre-training (Section 6.3.1). Initialization is from the mid-training T2I checkpoint (not the final aligned model), because its "unconstrained parameter space offers superior plasticity compared to post-trained models" for learning diverse editing tasks. Training follows a multi-scale strategy: 512×512 resolution with "massive, noisy datasets for rapid convergence," then 1024×1024 with high-quality data to refine details. Critically, joint training is employed — editing data is mixed with T2I mid-training data at a balanced batch ratio — to "improve both semantic understanding and image quality" and prevent catastrophic forgetting of generative knowledge. Additionally, each sample is associated with 3–5 candidate prompts (in Chinese and English), with one randomly selected per training iteration to enhance instruction generalization.

Stage 2: SFT (Section 6.3.2). A high-fidelity dataset of "hundreds of thousands of samples from real photographs, professional manual retouches, and synthetic sources" is curated with "rigorous human-in-the-loop filtering protocol, specifically targeting the structural alignment between source and edited images" (Section 6.3.2). The paper emphasizes high sensitivity to data quality: "even a marginal relaxation of these alignment standards leads to a precipitous drop in the model's ability to maintain consistency." Joint training with high-quality T2I SFT data continues from the pre-training stage, improving both instruction adherence and aesthetic quality.

Stage 3: DPO (Section 6.3.3). Described above in the RLHF subsection, using the preference dataset and DPO objective adapted for editing.

Why Separate Editing and T2I Models (Section 6.4). The paper initially aimed to unify T2I and editing into a single model but discovered a "critical data quality mismatch: the heavy reliance on synthetic data during editing pre-training noticeably degrades the photorealism of T2I generation compared to models trained solely on real data" (Section 6.4). The separate models are therefore "a data-driven issue, not an architectural flaw," and the paper believes that "by substituting synthetic datasets with large-scale interleaved corpora, future iterations can successfully merge these capabilities into a unified model without sacrificing generation quality" (Section 6.4).

4. Key Insights and Innovations

Innovation 1: AIGC-Contaminated Training Data as a Primary Bottleneck for Photorealism

The paper's most conceptually significant contribution is a diagnostic insight rather than a methodological one: trace amounts of AI-generated content in training data cause a qualitatively distinct failure mode — a "collapse" in the realistic generation subspace — that parameter scaling alone cannot resolve. This is not an incremental observation about data quality. It reframes the photorealism problem from one of model capacity (more parameters = more capacity to model complex textures) to one of data poisoning (synthetic images in the training distribution actively constrain the model away from photorealistic outputs).

What the field assumed before this work: The dominant hypothesis, implicit in the parameter-scaling trend the paper critiques (Section 1), was that photorealism is primarily limited by model expressivity — larger models with more parameters can capture finer texture details and more nuanced lighting, leading to more realistic outputs. The corollary was that training on larger datasets, regardless of provenance, would improve realism by providing more diverse examples. This assumption underlies the web-scale data collection strategies used by most large image generation models, where filtering focuses on resolution, aesthetics, and NSFW content but rarely on distinguishing human-captured from AI-generated images.

What the paper demonstrates instead: The paper's empirical finding, stated bluntly in Section 1, is that "even a small proportion of AIGC-contaminated data can cause the model to prematurely converge to a narrow local optimum during training." The mechanism (Section 2.1.4) is that AI-generated images share subtle but systematic statistical signatures — the "plastic" or "greasy" texture the paper repeatedly references — that, while individually imperceptible, collectively define an attractor basin in the model's generative manifold. When even a small fraction of training data comes from this basin, the model's optimization path is deflected toward it, and the model "collapses" away from the subspace of genuine photographic textures.

This is a fundamentally different diagnosis than "the model isn't big enough to capture realism." It implies that a 6B-parameter model trained on rigorously decontaminated data can achieve photorealism that eludes a 20B model trained on web-scale data with even modest AIGC contamination — not because the 6B model is more expressive, but because the 20B model's training signal is actively corrupted.

Significance beyond performance: This insight has two far-reaching implications. First, it explains a phenomenon many practitioners have observed but struggled to diagnose: the characteristic "AI-generated look" that persists even as models scale up. If AIGC contamination is the root cause, then no amount of parameter scaling will eliminate it — in fact, larger models trained on larger web-scale datasets may be more susceptible because the web's AIGC fraction is growing over time. Second, it shifts the research priority from architectural innovation (designing better transformer blocks) to data forensics (detecting and excluding AIGC content from training corpora). The paper's dual use of its AIGC detector — first as a data filter (Section 2.1.1), then as an adversarial reward model during RL (Section 4.3.2) — operationalizes this shift: the detector doesn't just exclude bad data, it actively penalizes the model for producing outputs that resemble that data.

Evidence: The paper does not provide a controlled ablation showing photorealism with vs. without AIGC filtering (this is a limitation of the report format — such ablations are implied by the design rationale but not explicitly reported). However, the qualitative results (Figures 15–18) and human evaluation (Figure 14, where LongCat-Image matches or exceeds Seedream 4.0 on Visual Realism despite being a fraction of its size) provide circumstantial evidence. The claim is further supported by the explicit use of the AIGC detector as a reward model during RL (Section 4.3.2), which would be unnecessary if AIGC contamination were not a significant limiting factor.

Innovation 2: Character-Level Tokenization as a Sufficient Alternative to Specialized Text Encoders

The paper's approach to Chinese text rendering represents a conceptual simplification that challenges a growing trend toward architectural complexity. Rather than designing a specialized text encoder for glyph rendering — the approach taken by GlyphByT5 (Liu et al., 2024), which the paper explicitly cites as an alternative — LongCat-Image achieves superior Chinese character coverage through a minimal intervention: switching from subword to character-level tokenization for quoted text spans in the existing LLM-based text encoder.

What the field assumed before this work: The dominant approach to improving text rendering in diffusion models has been to augment or replace the text encoder with specialized components. GlyphByT5 fine-tunes a ByT5 model specifically for glyph-aware encoding. Other works explore dedicated text renderer modules, separate text-to-glyph pathways, or post-processing correction networks. The implicit assumption is that text rendering requires specialized architectural support — that a general-purpose text encoder, even a powerful LLM, lacks the representational granularity to condition accurate glyph generation because its tokenization fragments characters across subword boundaries.

What the paper demonstrates instead: The character-level tokenization strategy (Section 3.2) is almost embarrassingly simple — identify quoted spans in the prompt, tokenize them at character granularity, and feed them through the same Qwen2.5-VL encoder. No new parameters. No specialized training. No architectural modification beyond the tokenizer configuration. Yet the results on ChineseWord (Table 7) are dramatic: 90.7% overall accuracy vs. 56.6% for Qwen-Image (which uses the same Qwen2.5-VL encoder but with standard subword tokenization) and 58.5% for the commercial Seedream 4.0. On Level 3 characters (the rarest), LongCat-Image achieves 70.3% vs. 6.1% and 2.3% respectively — roughly an order of magnitude improvement.

The insight is that the bottleneck was never the encoder's representational capacity. Qwen2.5-VL has more than enough capacity to encode glyph information — it is a 7B-parameter model trained on massive multimodal data. The bottleneck was the tokenization granularity: subword tokenization fragments rare Chinese characters across multiple tokens, creating a combinatorial explosion in the conditioning signal that makes learning glyph forms intractable without enormous data scale. Character-level tokenization collapses this fragmentation to a single token per character, reducing the learning problem from "compose a glyph from fragmented subword embeddings" to "learn a direct mapping from character embedding to glyph," which is feasible with modest synthetic data (10M+ SynthDoG samples).

Significance beyond performance: This finding has important implications for the design of multilingual generative models. It suggests that specialized text encoders (GlyphByT5, dedicated OCR-aware modules) may be unnecessary if the base encoder is sufficiently powerful and the tokenization strategy is adapted to the task. The computational savings are substantial: specialized encoders add parameters, memory footprint, and engineering complexity, while a tokenizer change costs essentially nothing. Moreover, the finding generalizes beyond Chinese — any language with a large character set and a long-tail distribution (Japanese kanji, Korean hanja, mathematical symbols, emoji) could potentially benefit from the same character-level tokenization approach, suggesting that the technique is not Chinese-specific but rather a general solution to the long-tail glyph rendering problem.

Distinction from incremental improvement: This is a fundamental conceptual simplification, not an incremental refinement. It collapses a problem that the field was addressing through architectural complexity (specialized encoders, dedicated modules) into a problem that can be addressed through data representation (how you tokenize the input). The 40+ percentage point gap on ChineseWord Level 3 characters between LongCat-Image and the next-best model is not a marginal gain — it represents the difference between functional and non-functional for rare character rendering.

Evidence: Table 7 (ChineseWord benchmark) and Tables 5–6 (GlyphDraw2 and CVTG-2K) provide the primary quantitative evidence. The qualitative comparisons in Figures 16–18 show that the character-level tokenization produces not just accurate individual characters but also natural text integration with complex backgrounds — a secondary benefit that the dynamic synthetic data sampling strategy (Section 4.1) explicitly targets by phasing out simplistic synthetic text data in favor of real-world text-in-image data once basic glyph forms are learned.

Innovation 3: Mid-Training as a Distinct and Releasable Model State

The paper introduces a conceptual distinction that does not exist in standard diffusion model training pipelines: mid-training as a qualitatively distinct phase that produces a model with different properties than either pre-trained or post-trained checkpoints — and one valuable enough to release as a separate artifact. This is not merely "more fine-tuning." It is a deliberate intervention that repositions the training pipeline from a linear progression (pre-train → align → deploy) to a branching tree where intermediate states are independently useful.

What the field assumed before this work: Standard diffusion model pipelines follow a two-phase structure: pre-training on broad data to learn the generative mapping, followed by some form of alignment (SFT, RLHF, or both) to refine quality and align with human preferences. Intermediate checkpoints between these phases are typically treated as transient — saved for resumption purposes but not considered independently valuable, because the assumption is that alignment strictly improves the model and any earlier checkpoint is simply an inferior version of the final model.

What the paper demonstrates instead: The mid-training checkpoint (Section 4.2) is:

  1. Qualitatively different from the pre-training checkpoint: it has been trained on curated high-fidelity data that constrains the learned manifold toward higher quality without the mode collapse risk of aggressive alignment.
  2. Qualitatively different from the post-training checkpoint: it "retains high plasticity and adaptability, avoiding the potential mode collapse or rigidity often introduced by aggressive alignment" (Section 4.2).
  3. Independently useful for specific downstream tasks: the image editing model is explicitly initialized from the mid-training checkpoint (not the final aligned model) because its "unconstrained parameter space offers superior plasticity compared to post-trained models" for learning diverse editing tasks (Section 6.3.1).

The insight is that alignment trades generality for specificity. SFT and RL narrow the model's generative distribution toward a particular aesthetic manifold — this improves average quality for standard generation tasks but reduces the model's ability to adapt to new tasks that require different output distributions. The mid-training checkpoint occupies a sweet spot: it has better quality than the pre-trained model (due to curated data) but broader capabilities than the aligned model (due to the absence of preference optimization).

Significance beyond performance: This reframes model release strategy. Rather than releasing a single "best" checkpoint, the paper argues for releasing the mid-training checkpoint as a "Developer Version" specifically designed for downstream fine-tuning. This is analogous to the distinction in LLMs between base models (pre-trained, suitable for fine-tuning) and instruct models (aligned, suitable for direct use), but applied at an intermediate granularity: the mid-training checkpoint is not a raw pre-trained model (it has received quality-focused training) but also not a fully aligned model (it has not been locked into specific aesthetic preferences). This tripartite release strategy (mid-training Developer Version + final aligned model + editing model) is unusual in the image generation field and reflects a deliberate philosophical choice about what kinds of artifacts best serve the research community.

Distinction from incremental improvement: This is a conceptual reframing with practical consequences for how models are released and used. The mid-training checkpoint is not just "the model before SFT" — it is specifically designed (through data curation choices in Section 2.1.4 and the explicit decision to release it) to serve a different purpose than the final model. The paper's editing results (Tables 9–11, where the editing model achieves SOTA among open-source models) provide strong evidence that mid-training initialization is superior to post-training initialization for editing — a finding that would not have been discovered without the conceptual distinction between mid-training and post-training as different model states with different properties.

Evidence: The editing benchmarks (Tables 9–11) indirectly validate the mid-training initialization choice, since the editing model's SOTA performance among open-source models depends on this design decision. The paper explicitly compares against Qwen-Image-Edit (which presumably uses a different initialization strategy) and finds LongCat-Image-Edit superior on CEdit-Bench (Table 9: 7.67 vs. 7.52 overall EN score), GEdit-Bench (Table 10: 7.64 vs. 7.56 overall EN score), and ImgEdit-Bench (Table 11: 4.50 vs. 4.27 overall score). The human evaluation (Figure 22) further supports the editing model's quality, showing win rates of 56.9% for comprehensive quality and 56.3% for consistency against Qwen-Image-Edit.

Innovation 4: Ensemble Reward Models Targeting Specific Failure Modes as a Substitute for Scale

The paper's RLHF strategy is distinguished not by algorithmic novelty — DPO and GRPO are established methods — but by the composition of its reward model ensemble, which directly targets the specific failure modes that parameter scaling fails to address. This is a conceptual shift from using RL to generally "improve quality" (the typical approach) to using RL as a targeted adversarial intervention against known degradation patterns.

What the field assumed before this work: The standard RLHF approach for image generation uses aesthetic preference models as reward signals — models trained to predict human quality judgments on dimensions like composition, lighting, and overall appeal. The implicit theory is that optimizing against a general preference model will push the generator toward higher-quality outputs across all dimensions simultaneously. This is a "raise all boats" strategy that assumes quality is roughly one-dimensional and that improvements are uniform.

What the paper demonstrates instead: The reward model ensemble (Section 4.3.2) includes four specialized detectors — AIGC detection, distortion detection, OCR accuracy, and human preference — each targeting a different and partially adversarial quality dimension. The critical insight is in the composition:

  • Human preference assessment pushes the model toward aesthetically pleasing outputs (the standard RL signal).
  • Distortion detection penalizes structural artifacts, anatomical errors, and unnatural proportions — failure modes that aesthetic models might not catch because they focus on global composition rather than local coherence.
  • OCR accuracy rewards correct text rendering — a capability that is orthogonal to photorealism and aesthetics, and one that standard preference models are not designed to evaluate.
  • AIGC detection is the most innovative component: it acts as an adversarial signal that specifically penalizes the "plastic" or "greasy" texture characteristic of AI-generated images, pushing the model away from the manifold of synthetic-looking outputs and toward genuine photographic textures.

The ensemble thus addresses failure modes that are non-overlapping and sometimes in tension. Improving OCR accuracy by rendering clearer text might, in a naïve optimization, reduce photorealism by producing text that looks pasted-on rather than naturally integrated. The ensemble resolves this tension by simultaneously rewarding text accuracy (OCR model) and photorealism (AIGC detector), forcing the model to find solutions that satisfy both constraints — text that is accurate and naturally integrated into the scene.

Significance beyond performance: This approach reframes RL for generative models from a monolithic "make it better" signal to a multi-objective optimization where different reward models provide complementary and sometimes competing gradients. It suggests that the key to effective RL is not a more sophisticated algorithm or a larger preference dataset, but rather a more diagnostic reward composition — identifying the specific failure modes of the base model and designing reward components that directly target each one. The AIGC detector as a reward model is particularly significant because it inverts the typical relationship between discriminators and generators: rather than training a discriminator to distinguish real from fake images (as in GANs), the paper uses a pre-trained AIGC detector to provide an anti-synthetic gradient, pushing the generator away from outputs that resemble the detector's training distribution.

Distinction from incremental improvement: The ensemble reward composition is a conceptual advance in what RL is optimizing for, not an algorithmic advance in how optimization is performed. The individual components (DPO, GRPO, MPO) are existing algorithms; the innovation is in the multi-faceted reward signal they optimize against. This is a fundamental shift from "RL for general quality improvement" to "RL for targeted failure mode correction," and it implies a different workflow for deploying RL in generative models: first diagnose the model's specific failure modes through careful evaluation, then design reward components that provide gradients against each failure mode, then combine them into an ensemble that balances competing objectives.

Evidence: The paper does not provide an ablation isolating the contribution of each reward model component (this is a significant limitation — we cannot determine from the report whether the AIGC detection reward, specifically, drives the photorealism improvements, or whether the OCR reward drives the text rendering gains). However, the human evaluation results (Figure 14) show LongCat-Image achieving a slight edge over the commercial Seedream 4.0 in Visual Realism, which is consistent with the AIGC-adversarial reward having a measurable effect. The ChineseWord benchmark (Table 7: 90.7% overall) provides circumstantial evidence for the OCR reward's contribution, since text rendering accuracy is a specific target of that reward component. The use of multiple RL stages (DPO for broad structural fixes, GRPO/MPO for fine-grained refinement) further supports the targeted-intervention framing: different RL algorithms are deployed against different aspects of the quality problem at different scales.

Innovation 5: Joint T2I + Editing Training as a Forgetting Prevention Strategy

The paper's approach to training the editing model introduces a simple but conceptually important technique: joint training with text-to-image data throughout the editing pre-training and SFT stages to prevent catastrophic forgetting of generative knowledge. This reframes the editing problem from one of task specialization (train an editor that forgets generation) to one of capability expansion (train a model that can do both, even if the final deployment separates them).

What the field assumed before this work: Standard practice for training specialized image models — including editing models — is to fine-tune a pre-trained generative model on task-specific data, accepting some degree of catastrophic forgetting as inevitable. The fine-tuned model becomes an "editing model" that may lose some or all of its text-to-image generation capability, but this is considered acceptable because the model is deployed only for editing. The assumption is that editing and generation are sufficiently different tasks that joint training would create interference — gradients from editing data would conflict with gradients from generation data, degrading both.

What the paper demonstrates instead: The joint training strategy (Sections 6.3.1 and 6.3.2) mixes editing data with T2I mid-training (or SFT) data at a balanced batch ratio during both the editing pre-training and SFT stages. The paper reports that "experiments confirm this approach improves both semantic understanding and image quality" (Section 6.3.1). The mechanism is straightforward: the T2I data provides a regularizing signal that keeps the model's generative manifold from collapsing toward the narrower editing distribution. The model learns to edit without unlearning how to generate.

The insight is that editing and generation are not competing objectives but mutually reinforcing capabilities. Understanding how to generate images from scratch (T2I) provides the model with knowledge about object structure, lighting, composition, and scene semantics that directly transfers to editing tasks (where the model must understand what to preserve and what to change). Conversely, learning to edit — which requires fine-grained understanding of image content and precise instruction following — may improve the model's ability to follow complex generation prompts.

Significance beyond performance: This finding challenges the assumption that task specialization requires sacrificing generality. It suggests that the catastrophic forgetting observed in standard fine-tuning is not an inevitable consequence of learning a new task but rather a symptom of insufficient regularization — the model forgets because the editing data distribution is too narrow compared to the generation data distribution, and joint training widens the effective distribution enough to preserve generality. The paper's eventual decision to separate the T2I and editing models (Section 6.4) was driven by a data quality mismatch (synthetic editing data degraded photorealism), not by a fundamental incompatibility of the tasks. The paper explicitly states this is "a data-driven issue, not an architectural flaw" and that "by substituting synthetic datasets with large-scale interleaved corpora, future iterations can successfully merge these capabilities into a unified model without sacrificing generation quality" (Section 6.4).

This has implications beyond editing: any specialized fine-tuning of generative models (inpainting, super-resolution, style transfer, subject-driven generation) could potentially benefit from joint training with general T2I data to preserve the base model's broad capabilities while adding specialized functionality.

Distinction from incremental improvement: Joint training itself is not novel — multi-task learning is a well-established technique. What is novel is the application to diffusion model editing and the explicit framing as a forgetting-prevention strategy rather than a performance-boosting strategy. The paper's ablation (implied by the "experiments confirm" language in Section 6.3.1, though not reported in detail) showing that joint training improves both semantic understanding and image quality suggests that the technique provides benefits beyond simply preventing forgetting — it actively improves editing performance by keeping the model grounded in generative knowledge.

Evidence: The editing benchmark results (Tables 9–11), combined with the qualitative editing examples (Figures 23–29), demonstrate that the editing model achieves SOTA open-source performance while maintaining consistency across complex multi-turn edits (Figure 23). The joint training contribution cannot be isolated from these results without a specific ablation, but the paper's explicit attribution of semantic understanding improvements to joint training (Section 6.3.1) indicates that the technique was empirically validated during development.

5. Experimental Analysis

Evaluation Methodology

  • Dataset. The paper evaluates text-to-image generation on five public benchmarks — GenEval (Ghosh et al., 2023) for fine-grained controllability, DPG-Bench (Hu et al., 2024) with 1,065 dense prompts for semantic alignment, WISE (Niu et al., 2025) with 1,000 curated prompts for world knowledge reasoning, GlyphDraw2 (Ma et al., 2025a) with 400 prompts across poster and complex subsets for bilingual text rendering, and CVTG-2K (Du et al., 2025) for multi-region English text rendering — plus a proprietary ChineseWord benchmark of 8,105 prompts covering the General Standard Chinese Characters Table, and internal Poster & Scene Text datasets (500 prompts) targeting business scenarios. Image editing is evaluated on CEdit-Bench (1,464 bilingual editing pairs across 15 task categories, introduced by the paper), GEdit-Bench (Liu et al., 2025) for Chinese and English instruction following, and ImgEdit-Bench (Ye et al., 2025) for multi-dimensional editing assessment. Human evaluation uses a custom dataset of 400 prompts spanning entity depiction to complex scene synthesis, with a separate 400-sample set for editing evaluation.

  • Base model(s). LongCat-Image is a 6B-parameter diffusion transformer built on the FLUX.1-dev architecture with a Qwen2.5VL-7B text encoder. The paper compares against a range of open-source models (PixArt-α at 0.6B, SD-3-Medium at 8B, FLUX.1-dev, SD-3.5-large, Janus-Pro-7B, HiDream-I1-Full, Lumina-Image 2.0, Qwen-Image at 20B, HunyuanImage-3.0 at 80B MoE) and commercial systems (Seedream 4.0, GPT Image 1, Nano Banana/Gemini-2.5-flash-image). For editing, baselines include FLUX.1 Kontext, Step1X-Edit, Qwen-Image-Edit, Seedream 4.0, Nano Banana, and GPT Image 1.

  • Metrics. GenEval reports accuracy across six sub-tasks (single object, two objects, counting, colors, position, attribute binding) with an overall average. DPG-Bench reports average accuracy across five dimensions (global, entity, attribute, relation, other). WISE reports category-level and overall accuracy. GlyphDraw2 reports per-subset and average accuracy. CVTG-2K reports word accuracy, Normalized Edit Distance (NED), and CLIPScore, broken down by number of text regions. ChineseWord uses PPOCRv5 for objective accuracy quantification across Level 1, Level 2, and Level 3 characters. Internal benchmarks report accuracy. Image editing uses Semantic Consistency (G_SC), Perceptual Quality (G_PQ), and Overall Score (G_O) evaluated by GPT-4o, consistent across CEdit-Bench, GEdit-Bench, and ImgEdit-Bench (which adds task-specific breakdowns). Human evaluation uses Mean Opinion Score (MOS) on four dimensions (text-image alignment, visual plausibility, visual realism, aesthetics) and Side-by-Side (SBS) win rates for editing.

  • Baselines. For text-to-image: Emu3, Show-o, PixArt-α/Σ, SD-3-Medium/Large, FLUX.1-dev, JanusFlow/Pro, Lumina-Image 2.0, HiDream-I1-Full, GPT Image 1, Seedream 3.0/4.0, Qwen-Image, HunyuanImage-3.0 (Tables 2–8). For text rendering specifically: Seedream 4.0, Qwen-Image, HunyuanImage-3.0 (Tables 5–8). For editing: FLUX.1 Kontext [Pro/Dev], GPT Image 1, Nano Banana, Seedream 4.0, Step1X-Edit, Qwen-Image-Edit (two versions), plus additional baselines on ImgEdit-Bench including MagicBrush, InstructPix2Pix, AnyEdit, UltraEdit, OmniGen/OmniGen2, ICEdIt, BAGEL, UniWorld-V1 (Tables 9–11). Human evaluation baselines: HunyuanImage-3.0, Qwen-Image, Seedream 4.0 for generation (Figure 14); Nano Banana, Seedream 4.0, FLUX.1 Kontext [Pro], Qwen-Image-Edit [2509] for editing (Figure 22).

  • Generation budget / compute accounting. The paper does not report generation budgets (number of inference steps, number of samples) for benchmark evaluation. The comparison is on final model outputs using unspecified inference configurations (presumably standard settings for each model). The efficiency argument is based on parameter count (6B vs. 20B–80B competitors), not on FLOPs-matched inference budgets. Training compute is reported in terms of steps per stage (Table 1: 900K + 300K + 200K pre-training, 70K mid-training, 20K SFT, 4K DPO, 300 GRPO/MPO) with batch sizes, but no total FLOPs or GPU-hours are provided. This is a significant gap: the paper claims efficiency advantages based on parameter count alone, without demonstrating that total training compute is lower than competitors or that inference FLOPs are reduced for equivalent quality.

  • Cross-validation / statistical protocol. None reported. Benchmark evaluations appear to be single-run results on the standard test sets without confidence intervals or statistical significance tests. The human evaluation MOS uses 400 prompts but does not report inter-annotator agreement, number of annotators, or confidence intervals. The SBS human evaluation for editing uses ~400 samples with win rates computed as (#Win + 0.5 × #Tie)/#Total, but annotator details are not provided. The ChineseWord benchmark uses PPOCRv5 for automated scoring, which introduces potential systematic bias (OCR models may have their own character-level accuracy limitations) that is not calibrated against ground-truth human verification.

Main Quantitative Results

Text-to-Image Alignment and General Generation

GenEval (Table 2). LongCat-Image achieves an overall score of 0.87, matching Qwen-Image (0.87) and exceeding all other open-source models: SD-3.5-large (0.71), FLUX.1-dev (0.66), Janus-Pro-7B (0.80), HiDream-I1-Full (0.83), and HunyuanImage-3.0 (0.72). It ties with Qwen-Image for the highest open-source score while using 3.3× fewer parameters (6B vs. 20B). Against commercial systems, it performs comparably to Seedream 4.0 (0.84), GPT Image 1 (0.84), and Seedream 3.0 (0.84). The sub-task breakdown reveals specific strengths and weaknesses: LongCat-Image scores 0.99 on Single Object (matching most competitors), 0.98 on Two Objects (matching HiDream-I1-Full's 0.98, the best open-source result), 0.86 on Counting (below Qwen-Image's 0.89 and Seedream 3.0's 0.91 but above most competitors), 0.86 on Colors (below several models including Seedream 3.0's 0.93 and Janus-Pro-7B's 0.90), 0.75 on Position (below Qwen-Image's 0.76 and Janus-Pro-7B's 0.79 but above FLUX.1-dev's 0.22), and 0.73 on Attribute Binding (below Qwen-Image's 0.77 and Seedream 4.0's 0.74). The pattern suggests LongCat-Image is strongest at entity presence and weakest at color specification and attribute binding — capabilities that may depend more on text encoder quality than on diffusion model scale.

DPG-Bench (Table 3). LongCat-Image achieves an overall score of 86.80, ranking below Qwen-Image (88.32), Seedream 4.0 (88.25), Seedream 3.0 (88.27), and Lumina-Image 2.0 (87.20), but above HunyuanImage-3.0 (86.10), GPT Image 1 (85.15), and all other open-source models. The dimensional breakdown shows: Global 89.10 (below Qwen-Image's 91.32 and Seedream 4.0's 94.10), Entity 92.54 (competitive with Qwen-Image's 91.56), Attribute 92.00 (competitive with Qwen-Image's 92.02 and above HunyuanImage-3.0's 89.13), Relation 93.28 (below Qwen-Image's 94.31 and Lumina-Image 2.0's 94.85), Other 87.50 (below Qwen-Image's 92.73 and Seedream 4.0's 92.77). The gap is largest on "Other" (miscellaneous semantic aspects), suggesting LongCat-Image's more compact architecture may sacrifice some breadth of semantic coverage compared to the 20B Qwen-Image.

WISE (Table 4). LongCat-Image achieves an overall score of 0.65, ranking second among text-to-image models behind Seedream 4.0 (0.78) and ahead of Qwen-Image (0.62), HunyuanImage-3.0 (0.57), and FLUX.1-dev (0.50). The dimensional breakdown: Cultural 0.66 (vs. Seedream 4.0's 0.78), Time 0.61 (vs. 0.73), Space 0.72 (vs. 0.85), Biology 0.66 (vs. 0.79), Physics 0.72 (vs. 0.84), Chemistry 0.49 (vs. 0.67). The consistent gap with Seedream 4.0 across all dimensions is notable but must be contextualized against the massive parameter and data advantage of the commercial system. Against unified models (GPT-4o at 0.80, MetaQuery-XL at 0.55), T2I models generally underperform, and LongCat-Image's score of 0.65 represents the highest T2I result except Seedream 4.0.

Text Rendering

GlyphDraw2 (Table 5). LongCat-Image achieves an average score of 0.95 across the four subsets (Complex-en 0.94, Complex-zh 0.92, Poster-en 0.95, Poster-zh 0.99), ranking between Seedream 4.0 (0.97 average) and Qwen-Image (0.93 average), and substantially above HunyuanImage-3.0 (0.78 average). On Complex-zh — the most challenging subset testing random combinations of 2,000 frequent Chinese characters — LongCat-Image's 0.92 narrowly exceeds Seedream 4.0's 0.91 and Qwen-Image's 0.87, suggesting the character-level tokenization strategy provides particular benefits for complex character sequences. On Poster-zh (design-focused Chinese text), LongCat-Image achieves 0.99, matching Seedream 4.0 and slightly exceeding Qwen-Image's 0.98.

CVTG-2K (Table 6). LongCat-Image achieves the highest average Word Accuracy of 0.8658, substantially ahead of Qwen-Image (0.8288), Seedream 4.0 (0.8917 — noting that Seedream leads overall but the average here includes it), and HunyuanImage-3.0 (0.7650). Broken down by number of text regions, LongCat-Image leads at 2 regions (0.9129 vs. Qwen-Image's 0.8370, Seedream 4.0's 0.8898) and trails Seedream 4.0 at 3, 4, and 5 regions, with the gap widening as complexity increases (0.8310 vs. 0.8873 at 5 regions). NED follows the same pattern (0.9361 overall, below Seedream 4.0's 0.9507 but above Qwen-Image's 0.9297). CLIPScore is similar across models (LongCat-Image 0.7859, Qwen-Image 0.8059, Seedream 4.0 0.7853). This suggests LongCat-Image's text rendering is strongest for simpler layouts and degrades more noticeably than commercial systems as the number of text regions increases — consistent with the paper's own acknowledgment that "the model experiences a noticeable decline in stability when generating multi-character sequences" (Section 5.1.2).

ChineseWord (Table 7). This is LongCat-Image's most dramatic result. Overall accuracy is 90.7%, compared to Seedream 4.0's 58.5%, Qwen-Image's 56.6%, and HunyuanImage-3.0's 49.3%. The breakdown by character frequency tier reveals where the gap originates: Level 1 (most common characters) — LongCat-Image 98.7% vs. Seedream 4.0 94.8%, Qwen-Image 92.5%, HunyuanImage-3.0 83.5%. The real separation occurs at Level 2: LongCat-Image 90.8% vs. Seedream 4.0 41.2%, Qwen-Image 37.1%, HunyuanImage-3.0 31.3%. At Level 3 (rarest characters), the gap becomes enormous: LongCat-Image 70.3% vs. Seedream 4.0 2.3%, Qwen-Image 6.1%, HunyuanImage-3.0 4.1%. This 68 percentage point gap at Level 3 is not marginal — it represents the difference between a model that can render rare characters most of the time and models that essentially cannot render them at all. The dramatic drop from Level 2 to Level 3 for competitors (41.2% → 2.3% for Seedream 4.0) versus LongCat-Image's more gradual decline (90.8% → 70.3%) directly supports the paper's claim that character-level tokenization and synthetic text data training address the long-tail character distribution in a way that scaling alone cannot.

Poster & Scene Text (Table 8). On internal benchmarks targeting production scenarios, LongCat-Image achieves 91.5% average accuracy (Poster 92.0%, Real Scene 91.0%), essentially matching Seedream 4.0's 91.6% and exceeding Qwen-Image's 89.2% and HunyuanImage-3.0's 87.1%. On Real Scene text specifically, LongCat-Image's 91.0% slightly exceeds Seedream 4.0's 90.0%, suggesting the model's text integration into natural environments is competitive with commercial systems.

Human Evaluation for Text-to-Image

MOS Results (Figure 14). The human evaluation on four dimensions reveals a nuanced picture. LongCat-Image demonstrates "comprehensive superiority over HunyuanImage 3.0 across all metrics" (Section 5.2). Compared to Qwen-Image, LongCat-Image achieves parity in Text-Image Alignment and Visual Plausibility. On Visual Realism, LongCat-Image "excels... outperforming Qwen-Image and even exhibiting a slight advantage over the commercial baseline, Seedream 4.0" (Section 5.2). On Visual Aesthetics, there remains "a marginal gap in Visual Aesthetics compared to Qwen-Image." The paper does not provide the numerical MOS scores — only the directional comparisons are reported, which limits quantitative analysis.

Image Editing

CEdit-Bench (Table 9). Among open-source models, LongCat-Image-Edit achieves the highest scores on both English (G_SC 8.27, G_PQ 7.88, G_O 7.67) and Chinese (G_SC 8.25, G_PQ 7.85, G_O 7.65). This places it ahead of Qwen-Image-Edit (EN: 8.07/7.84/7.52; CN: 8.03/7.78/7.46) and Qwen-Image-Edit [2509] (EN: 8.04/7.79/7.48; CN: 7.93/7.71/7.37), as well as FLUX.1 Kontext [Dev] (EN: 6.31/7.56/5.93) and Step1X-Edit (EN: 6.68/7.36/6.25). Against commercial systems, LongCat-Image-Edit trails GPT Image 1 (EN: 8.64/8.26/8.17), Seedream 4.0 (EN: 8.12/7.95/7.58), and Nano Banana (EN: 7.51/8.17/7.20) on English, though the gap with Seedream 4.0 is modest (7.67 vs. 7.58 Overall) and the gap with Nano Banana on Chinese is small (7.65 vs. 7.36). A notable data point: FLUX.1 Kontext [Pro] scores dramatically lower on Chinese (G_SC 1.15, G_PQ 8.07, G_O 1.43) than on English (6.79/7.80/6.53), indicating near-total failure on Chinese editing instructions — a limitation LongCat-Image-Edit does not share (CN scores are comparable to EN scores).

GEdit-Bench (Table 10). On English, LongCat-Image-Edit achieves G_SC 8.18, G_PQ 8.00, G_O 7.64, placing it slightly above Qwen-Image-Edit [2509] (8.15/7.86/7.54) and Qwen-Image-Edit (8.00/7.86/7.56), and below Seedream 4.0 (8.24/8.08/7.68) and Nano Banana (7.86/8.33/7.54). On Chinese, LongCat-Image-Edit scores 8.08/7.99/7.60, compared to Qwen-Image-Edit [2509] (8.05/7.88/7.49) and Seedream 4.0 (8.19/8.14/7.71). The pattern is consistent: LongCat-Image-Edit leads open-source models by a small but consistent margin, trails commercial systems by a similarly small margin, and shows essentially no performance degradation between English and Chinese (unlike several baselines).

ImgEdit-Bench (Table 11). LongCat-Image-Edit achieves the highest overall score of 4.50, surpassing all competitors including commercial systems: Nano Banana (4.35), Seedream 4.0 (4.18), GPT Image 1 (4.20), and FLUX.1 Kontext [Pro] (4.00). The task-level breakdown shows LongCat-Image-Edit leading or tying for the lead on seven of nine tasks: Add (4.51, second to GPT Image 1's 4.61), Adjust (4.57, the highest), Extract (3.93, second to Qwen-Image-Edit [2509]'s 4.04), Replace (4.76, the highest), Remove (4.60, the highest), Background (4.49, second to Seedream 4.0's 4.30? — actually LongCat leads here at 4.49 vs. Qwen-Image-Edit [2509] at 4.37), Style (4.85, the highest), Hybrid (4.01, the highest), and Action (4.74, second to GPT Image 1's 4.89). This is a notably strong result suggesting that on this particular benchmark's task composition and metric, the editing model's performance is genuinely competitive with or superior to commercial alternatives.

Human Evaluation for Editing

SBS Results (Figure 22). Against Qwen-Image-Edit [2509], LongCat-Image-Edit achieves win rates of 56.9% for Comprehensive Quality and 56.3% for Consistency. Against FLUX.1 Kontext [Pro], the win rates are 60.5% and 60.8% respectively. However, against commercial systems, the model trails: against Nano Banana, win rates are 43.1% (Comprehensive Quality) and 39.5% (Consistency); against Seedream 4.0, 43.7% and 37.0% respectively. This quantifies the paper's own assessment that "a performance gap remains when compared to commercial systems such as Nano Banana and Seedream 4.0" (Section 6.5.2).

Ablation Studies and Robustness Checks

AIGC detection as reward model (implied ablation). The paper does not provide an explicit ablation isolating the AIGC detection reward model's contribution to photorealism. The human evaluation result showing LongCat-Image with a slight edge over Seedream 4.0 on Visual Realism (Figure 14) is the primary evidence for the AIGC-adversarial training's effectiveness, but without a controlled comparison (same model trained with and without the AIGC reward), the causal claim cannot be verified from this report alone.

Character-level tokenization (implied ablation). No explicit ablation compares LongCat-Image's character-level tokenization against standard subword tokenization while holding all other factors constant. The ChineseWord benchmark results (Table 7) provide strong circumstantial evidence — LongCat-Image's 90.7% vs. Qwen-Image's 56.6% using the same Qwen2.5-VL encoder — but these models differ in many other respects (architecture, training data, training procedure), so the tokenizer's specific contribution is confounded.

Multi-granularity captioning (implied ablation). The paper describes the weighted sampling strategy for four caption granularities (Section 2.1.3) but does not report an ablation comparing this against single-granularity captioning or alternative mixing ratios. The design rationale is clearly articulated but empirically unvalidated in the reported results.

Mid-training vs. post-training initialization for editing. The editing model's SOTA open-source performance (Tables 9–11) supports the claim that mid-training initialization is beneficial, but no direct ablation compares editing performance when initialized from the mid-training checkpoint versus the post-training (SFT/RL) checkpoint. The paper's statement is based on internal experiments referenced as "experiments confirm" (Section 6.3.1) without providing the comparative data.

Joint T2I + editing training. Similarly, the paper claims that joint training "improves both semantic understanding and image quality" (Section 6.3.1) and "effectively mitigates catastrophic forgetting" (Section 1), but no ablation isolates this effect. The editing benchmarks show strong results but cannot disentangle the joint training contribution from the other design choices (mid-training initialization, data quality filtering, DPO).

Synthetic text data dynamic sampling (Section 4.1). The dynamic strategy — increasing sampling probability for high-error characters, phasing out synthetic data in final pre-training — is described in detail but no ablation compares it against static sampling or no phase-out. The ChineseWord and GlyphDraw2 results demonstrate the end-state text rendering quality but do not reveal the marginal contribution of the dynamic schedule.

GRPO vs. MPO comparison. The paper presents MPO as an improvement over GRPO, citing "superior training efficiency and stability" (Section 4.3.2), but provides no head-to-head experimental comparison of the two methods. The training configuration (Table 1) shows both use 300 steps, but no metrics (reward curves, final model quality, training time) are reported to substantiate the claimed advantages.

Model weight averaging for SFT. The paper merges multiple specialized SFT models via weight averaging (Section 4.3.1) but does not compare the merged model against individual specialized models or against a single model trained on the combined dataset. The claim that merging "effectively balances performance across multiple attributes" (Section 4.3.1) is plausible but unvalidated.

Reward model ensemble composition. No ablation studies examine the contribution of individual reward models (AIGC detection, OCR accuracy, distortion detection, human preference) to final model quality. Training runs with subsets of the reward ensemble would reveal which reward components drive which quality improvements and whether any components are redundant or antagonistic.

Progressive resolution curriculum. The three-stage pre-training progression (256 → 512 → 512-1024px) is compared against a hypothetical direct 256 → 1024px jump (which the paper claims causes "computational instability," Section 4.1), but no empirical comparison is provided to quantify the benefit of the intermediate stage.

Critical Assessment

Claim 1: "With only 6B parameters, LongCat-Image surpasses numerous open-source models that are several times larger across multiple benchmarks." This claim is supported with qualifications. The benchmark results (Tables 2–8) consistently show LongCat-Image outperforming HunyuanImage-3.0 (80B MoE, ~13× larger) and achieving parity with or exceeding Qwen-Image (20B, 3.3× larger) on most metrics. However, the claim of "surpassing" requires specifying which benchmarks and which competitors. On GenEval, LongCat-Image ties Qwen-Image (0.87 vs. 0.87), which is parity, not superiority. On DPG-Bench, it trails Qwen-Image (86.80 vs. 88.32). On WISE, it leads Qwen-Image (0.65 vs. 0.62). On text rendering, it dramatically exceeds Qwen-Image on ChineseWord (90.7% vs. 56.6%) and GlyphDraw2 (0.95 vs. 0.93). The claim holds most strongly for Chinese text rendering and holds conditionally for general generation, where the advantage over Qwen-Image is benchmark-dependent. A critical caveat: total training compute is not reported, so we cannot determine whether LongCat-Image's efficiency advantage extends to training cost or only to inference-time parameter count.

Claim 2: "Through an innovative data strategy and training framework, our model achieves remarkable photorealism." This claim is supported by human evaluation but lacks controlled evidence. The MOS results (Figure 14) show LongCat-Image with a slight advantage over Seedream 4.0 on Visual Realism, which is impressive for an open-source 6B model. However, the causal attribution to specific data strategies (AIGC exclusion, AIGC reward model, mid-training quality curation) cannot be verified without ablations. The claim that AIGC-contaminated data causes "plastic" or "greasy" textures is an important diagnostic hypothesis, but the paper does not provide the controlled experiment that would confirm it: training otherwise identical models with and without AIGC filtering and measuring photorealism. The qualitative results (Figures 15–18) are visually compelling but subject to selection bias.

Claim 3: "The model demonstrates superior accuracy and stability in rendering common Chinese characters compared to existing SOTA open-source models and achieves industry-leading coverage of the Chinese dictionary." This claim is strongly supported by the ChineseWord benchmark (Table 7), which is the most rigorous and differentiating evaluation in the paper. The gap between LongCat-Image (90.7%) and the next-best open-source model (Qwen-Image at 56.6%) is enormous and consistent across all three difficulty levels, with the gap widening at rarer characters (Level 3: 70.3% vs. 6.1%). The benchmark design — 8,105 prompts covering the standard character table with automated OCR evaluation — is systematic and reproducible. However, two caveats: PPOCRv5, the evaluation OCR model, has its own accuracy limitations that are not calibrated, and single-character evaluation may not fully represent multi-character rendering quality (which the paper acknowledges degrades for all models). The "industry-leading coverage" claim is specifically about Chinese dictionary coverage and is well-supported; it should not be misinterpreted as claiming industry-leading text rendering in all contexts (the CVTG-2K English results show trailing Seedream 4.0).

Claim 4: "The LongCat-Image editing model achieves state-of-the-art performance among open-source models, delivering a leading performance of instruction following and image quality, as well as superior visual consistency." This claim is supported across multiple benchmarks. On CEdit-Bench, GEdit-Bench, and ImgEdit-Bench (Tables 9–11), LongCat-Image-Edit consistently leads all open-source competitors. The human evaluation SBS results (Figure 22) further support the claim against the strongest open-source baseline (Qwen-Image-Edit [2509]) with win rates of ~57% on both comprehensive quality and consistency. However, "state-of-the-art among open-source models" is a narrower claim than "state-of-the-art" — the model still trails commercial systems (Nano Banana, Seedream 4.0) on most metrics, sometimes substantially (Figure 22 shows ~43% win rate vs. Seedream 4.0). The "superior visual consistency" claim is specifically supported by the SBS consistency evaluation (56.3% vs. Qwen-Image-Edit) and by the qualitative multi-turn editing examples (Figure 23), but the gap is modest and the claim should be understood as "superior to other open-source models" rather than absolute superiority.

Claim 5: "We provide a complete toolchain, from intermediate checkpoints to the full training code." This claim is factual about release scope but cannot be evaluated for completeness from the paper alone. The release of mid-training checkpoints and training code is described (Section 4.2, Section 8), and the associated URLs are provided. Whether the released toolchain is "complete" enough to reproduce results is an empirical question for the community.

Genuine weaknesses in the experimental design:

  1. Missing ablations for nearly every claimed innovation. The paper's contributions — AIGC exclusion, character-level tokenization, multi-granularity captioning, mid-training initialization for editing, joint T2I+editing training, ensemble reward models, dynamic synthetic data sampling — are individually plausible but none are empirically isolated. The results demonstrate that the complete system works well, but cannot attribute performance to specific components. A reader cannot determine whether AIGC filtering is genuinely critical or whether the other data curation steps would suffice; whether character-level tokenization drives the ChineseWord results or whether the 10M+ synthetic text samples alone would achieve similar gains with standard tokenization; whether model weight averaging provides benefits over single-model SFT. This pattern is common in large-scale industry technical reports but limits the paper's scientific contribution: it demonstrates that a combination of techniques works, not why each technique matters or how much it contributes.

  2. No FLOPs or compute comparison. The paper's central thesis is that efficient design can substitute for parameter scaling, but training compute is never compared between LongCat-Image and its competitors. The 6B vs. 20B vs. 80B comparison is about inference-time parameters, not total training cost. A larger model trained for fewer steps might have lower total training FLOPs than a smaller model trained for more steps. Without training FLOPs data, the efficiency claim is incomplete — we know LongCat-Image is smaller at deployment, but not whether it is cheaper to train.

  3. Single model architecture (FLUX.1-dev variant). All experiments use one architecture family, and the paper's conclusions about the 6B parameter scale being "ideal" are based on ablation within this family only. Whether a different architecture at 6B (e.g., a pure DiT without MM-DiT blocks, a U-Net-based diffusion model, an autoregressive model) would achieve similar results is untested. The claim that 6B is the optimal scale is architecture-contingent.

  4. Human evaluation methodology is under-specified. The number of annotators, inter-annotator agreement metrics, annotation interface, and quality control procedures are not described. MOS scores are reported only directionally (Figure 14 shows bar charts without numerical labels). The SBS evaluation (Figure 22) reports win rates but not statistical significance. These omissions make it difficult to assess the reliability of the human evaluation results.

  5. ChineseWord benchmark uses PPOCRv5 as ground truth proxy. The OCR model's own accuracy on rare characters is not reported or calibrated against human annotation. If PPOCRv5 has systematic weaknesses on certain character types, the ChineseWord scores would reflect OCR bias rather than true rendering quality. This could inflate or deflate the reported accuracy in unknown ways.

  6. Test set sizes are small for some benchmarks. GenEval and DPG-Bench have established test sets, but the internal benchmarks (Poster & Scene Text: 500 prompts, ChineseWord: 8,105 prompts) and CEdit-Bench (1,464 pairs) are custom-created. The paper does not report how these were constructed, whether they overlap with training data, or whether they are representative of real-world distributions. The risk of inadvertent training set contamination is elevated for custom benchmarks.

  7. No latency or memory benchmarks. The paper claims "minimal VRAM usage and rapid inference" (Abstract) as an advantage of the 6B parameter scale but reports no VRAM measurements, no inference time comparisons, and no throughput benchmarks. A 6B model is smaller than a 20B model, but actual inference speed depends on implementation, quantization, batch size, and sampling steps — none of which are reported.

Experiments that would have strengthened the paper:

  1. Ablation of AIGC filtering and AIGC reward model: Train a 6B model identical to LongCat-Image but without AIGC exclusion from pre-training data and without the AIGC reward model during RL, then compare photorealism metrics and human evaluation scores. This would directly test the paper's central diagnostic claim.

  2. Ablation of character-level tokenization: Train a model identical to LongCat-Image but with standard subword tokenization (keeping the synthetic text data and training procedure otherwise identical), then compare ChineseWord accuracy. This would isolate the tokenizer's contribution.

  3. Component-wise reward model ablation: Train RL variants with different subsets of the reward ensemble (AIGC only, OCR only, preference only, combinations) to determine which components drive which quality dimensions.

  4. Mid-training vs. post-training editing initialization: Fine-tune two editing models — one from the mid-training checkpoint, one from the post-training (after SFT/RL) checkpoint — with identical editing data and training procedure, then compare editing benchmark performance.

  5. Training FLOPs comparison: Estimate total training FLOPs for LongCat-Image and for Qwen-Image (or another open-source competitor with reported training details), providing a more complete efficiency picture.

  6. Inference latency and memory benchmarks: Report generation time, VRAM usage, and throughput at standard resolutions for LongCat-Image and competitors, ideally on identical hardware.

  7. Confidence intervals on benchmark scores: For benchmarks with small test sets, report bootstrap confidence intervals to quantify the uncertainty in the reported comparisons, particularly where margins are small (e.g., LongCat-Image vs. Qwen-Image on GenEval at 0.87 vs. 0.87).

Conditional nature of the claims: The efficiency claim (6B matching or exceeding larger models) holds for Chinese text rendering unconditionally — the ChineseWord gap is too large to be explained by evaluation noise. For general T2I generation, it holds conditionally: LongCat-Image achieves parity with Qwen-Image on some benchmarks (GenEval, human evaluation alignment/plausibility) but trails on others (DPG-Bench, WISE compared to Seedream 4.0). The photorealism claim holds in human evaluation but is not isolated to specific interventions. The editing claim holds among open-source models but not against commercial systems. These conditional boundaries are not clearly demarcated in the paper's abstract or conclusion, which frame the results as broadly competitive, but the detailed benchmark tables provide the evidence for a more nuanced interpretation.

6. Limitations and Trade-offs

6.1 The Difficulty Estimation Cost Is Not Accounted for in the Efficiency Claims

The Assumption or Constraint: The paper's central efficiency claim — that a 6B model can match or exceed 20–80B parameter models — is based entirely on parameter count comparisons. However, the paper never accounts for the training compute cost required to achieve this parameter efficiency. The training recipe (Table 1) spans 1.4 million pre-training steps across three resolution stages, 70K mid-training steps, 20K SFT steps, 4K DPO steps, and 300 GRPO/MPO steps — all at batch sizes ranging from 32 to 4608. The total FLOPs or GPU-hours for this pipeline are not reported, nor are they compared against the training costs of competitors like Qwen-Image (20B) or HunyuanImage-3.0 (80B MoE). The abstract states the model is "significantly smaller than the nearly 20B or larger" architectures, and Section 5.1 frames all comparisons around parameter count, but parameter count at inference time is only one dimension of efficiency.

The Consequence: The headline efficiency claim is incomplete and potentially misleading in one direction. A 6B model trained for 1.4 million steps at batch size 4608 processes far more tokens during training than a 20B model trained for fewer steps at smaller batch size. Without training FLOPs data, we cannot determine whether LongCat-Image is genuinely more compute-efficient overall (training + inference combined), or whether it achieves inference-time efficiency at the expense of higher training-time cost. This matters for two practical scenarios:

  • Organizations that train models from scratch: If LongCat-Image requires 2× the training FLOPs of a 20B competitor to reach competitive quality, the total cost of ownership may favor the larger model, particularly if inference volume is high.
  • Reproducibility and research access: If the training compute is enormous (as the 4608 batch size and 1.4M steps suggest), the paper's stated goal of "democratizing the technology" (Section 1) is undermined — a 6B model that costs as much to train as a 20B model does not lower the barrier to entry for training, even if it lowers the barrier for deployment.

A secondary consequence: inference latency and memory are not benchmarked. The paper claims "minimal VRAM usage and rapid inference" (Abstract) but provides no measurements. A 6B diffusion model running at 1024px resolution with 30 DiT blocks and a 7B text encoder requires substantial VRAM — likely 16–24GB depending on precision — and inference speed depends on the number of sampling steps, which is unspecified. The claim of deployment efficiency is asserted, not demonstrated.

What Evidence Exists in the Paper: The paper reports training hyperparameters in Table 1 (steps, batch sizes, learning rates) but does not convert these to total FLOPs, GPU-hours, or estimated training cost. No inference benchmarks (VRAM, throughput, latency) are provided. The efficiency argument rests entirely on the 6B vs. 20B vs. 80B parameter count comparison.

Mitigation Status: Not addressed. The paper does not acknowledge this gap. The efficiency claims are presented as self-evident from parameter count alone, with no discussion of training cost amortization or inference benchmarking.

6.2 Hard Problems and Rare-Character Multi-Character Rendering Show Fundamental Capability Bounds

The Assumption or Constraint: The paper implicitly assumes that careful data curation and targeted training can close the gap between a 6B model and commercial-scale systems. However, the results reveal sharp capability boundaries where this assumption breaks down. The most important is acknowledged in Section 5.1.2:

"we acknowledge that while exhibiting dominance in single-character rendering, the model experiences a noticeable decline in stability when generating multi-character sequences, primarily due to the insufficient scale of real-world textual training data."

This is not a marginal failure. The CVTG-2K benchmark (Table 6) shows that LongCat-Image's word accuracy degrades from 0.9129 at 2 regions to 0.8310 at 5 regions, a drop of 8.2 percentage points that is steeper than Seedream 4.0's decline (0.8898 → 0.8873, a 0.3 point drop). The model leads at simple layouts but trails at complex multi-text layouts. Similarly, the hardest difficulty bin for general T2I benchmarks shows that the efficiency advantages narrow or reverse as task complexity increases — on DPG-Bench (Table 3), LongCat-Image trails Qwen-Image by 1.52 percentage points overall, with the gap concentrated in the "Other" dimension (87.50 vs. 92.73).

The Consequence: These capability boundaries mean that the 6B model's advantages are conditional, not universal. For use cases that require:

  • Rendering multiple distinct text elements in different regions of the same image (e.g., complex posters, infographics, multi-panel designs)
  • Handling the "Other" semantic category on DPG-Bench (miscellaneous complex semantic relationships)
  • Rivaling commercial systems on the hardest editing scenarios (Figure 22 shows 37–43% win rates vs. Seedream 4.0 and Nano Banana)

the model underperforms alternatives that use 3–13× more parameters. A practitioner deciding between deploying LongCat-Image or a larger model must understand that the efficiency gains come at the cost of degraded performance on the hardest sub-problems. The paper frames its results as "competitive" or "surpassing" without consistently demarcating where the model leads versus where it trails.

What Evidence Exists in the Paper: CVTG-2K (Table 6) shows the multi-region degradation. DPG-Bench (Table 3) shows the semantic breadth gap. The editing SBS evaluation (Figure 22) quantifies the gap against commercial systems. The multi-character sequence limitation is explicitly acknowledged in Section 5.1.2.

Mitigation Status: Acknowledged for multi-character text rendering only. The paper states that "in future work, we aim to address this by rigorously expanding our text-rich dataset collection to enhance robustness in complex multi-character generation tasks" (Section 5.1.2). The broader capability boundaries on semantic complexity and hardest editing tasks are not explicitly discussed as limitations. The abstract and conclusion frame the results as broadly competitive without qualification.

6.3 No Ablations Isolate the Contribution of Individual Design Choices

The Assumption or Constraint: The paper introduces at least seven distinct innovations — AIGC filtering from pre-training data, AIGC detection as an RL reward model, character-level tokenization for text rendering, multi-granularity captioning, mid-training initialization for the editing model, joint T2I + editing training, and ensemble reward models for RL — but provides zero controlled ablation experiments that isolate the contribution of any single component. Every result reflects the complete system with all innovations applied simultaneously. The paper frequently uses language like "experiments confirm" (Section 6.3.1) or "empirical evidence suggests" (Section 3.2) to describe design decisions, but these internal validations are not reported.

The Consequence: This is the most significant methodological weakness in the paper. Without ablations:

  • The central diagnostic claim — that AIGC contamination causes photorealism degradation — is unverified. The paper asserts this as a causal mechanism (Section 1: "even a small proportion of AIGC-contaminated data can cause the model to prematurely converge to a narrow local optimum") but never demonstrates it by training an otherwise-identical model without AIGC filtering and measuring the photorealism difference. The observed photorealism could be due to the curated mid-training data, the RL reward ensemble, the Qwen2.5-VL encoder's quality, or any combination of factors.
  • The contribution of character-level tokenization cannot be separated from the synthetic text data. The ChineseWord benchmark (Table 7) shows LongCat-Image dramatically outperforming Qwen-Image (90.7% vs. 56.6%), but both models differ in tokenization strategy, training data composition (10M+ synthetic text samples), and training procedure. We cannot determine whether character-level tokenization is the critical factor, or whether Qwen-Image with the same 10M synthetic text samples would achieve similar accuracy with subword tokenization.
  • The claimed benefits of model weight averaging, joint editing training, and MPO over GRPO are asserted without evidence. The paper states that MPO achieves "superior training efficiency and stability" (Section 4.3.2) but provides no training curves, reward trajectories, or comparative metrics. The model weight averaging strategy (Section 4.3.1) is described as "effectively balancing performance across multiple attributes" but no comparison against a single-model SFT baseline is reported.

The reproduction and adoption barrier is severe. A practitioner who wants to apply these techniques to a different model or domain cannot determine which components are essential and which are incidental. If training compute is limited, should they prioritize AIGC filtering from pre-training data, or the AIGC reward model during RL, or both? If they only care about English text rendering, is character-level tokenization still beneficial? The paper provides no guidance because it provides no decomposition of its results.

What Evidence Exists in the Paper: None. All results are for the complete integrated system. No ablation tables, no component-wise contributions, no controlled comparisons of "LongCat-Image without X" versus "LongCat-Image with X" are reported.

Mitigation Status: Not addressed. The paper presents the integrated system as a unified design philosophy ("Building efficient and powerful model," Section 1) and does not discuss the absence of ablations. This is typical of large-scale industry technical reports, where the goal is to describe a production system rather than to provide scientific decomposition, but it fundamentally limits the paper's value as a research contribution — it demonstrates that the combination works, not why or which parts matter.

6.4 Single Model Architecture Family and Single Core Benchmark Domain

The Assumption or Constraint: All experiments use a single model architecture family — the FLUX.1-dev hybrid MM-DiT / Single-DiT transformer with a Qwen2.5-VL encoder — and a single core benchmark domain — the MATH-analogous image generation benchmarks (GenEval, DPG-Bench, WISE, text rendering benchmarks, editing benchmarks). The paper claims that 6B parameters "serves as the ideal foundation for balancing capability and efficiency" (Section 1) based on "systematic experimentation," but this experimentation is confined to one architecture. The paper does not test:

  • Whether a different architecture at 6B (pure DiT without MM-DiT blocks, U-Net-based diffusion, autoregressive image generation) would achieve similar or better results
  • Whether the claimed advantages hold for domains beyond photorealistic generation and editing (e.g., medical imaging, satellite imagery, scientific visualization, 3D-aware generation)
  • Whether the character-level tokenization strategy transfers to other languages with large character sets (Japanese kanji, Korean hanja) or to other text-conditioned generation tasks

The Consequence: The claim that 6B is the "ideal" parameter scale is architecture-contingent. The paper has demonstrated that a well-designed 6B FLUX.1-dev variant with careful data curation can compete with 20B+ models — but it has not demonstrated that 6B is the optimal scale in general, or that the specific architectural choices (10 MM-DiT + 20 Single-DiT blocks, 3D M-RoPE, character-level tokenization with Qwen2.5-VL) are necessary rather than sufficient for the observed performance. A different architecture might achieve the same quality at 4B parameters, or might require 10B to match LongCat-Image's results. The efficiency claims are relative to specific competitors (Qwen-Image, HunyuanImage-3.0) with different architectures, not relative to a Pareto frontier of quality vs. parameter count across architectures.

For practitioners, the practical implication is uncertainty about transfer: if a team is using a U-Net-based diffusion model or an autoregressive architecture, the paper provides no evidence about whether the data curation and training strategies would yield similar benefits. The findings are tightly coupled to the FLUX.1-dev transformer architecture.

What Evidence Exists in the Paper: All benchmarks in Tables 2–11 use LongCat-Image in its single architectural configuration. No architectural variants are compared. The systematic experimentation that determined 6B as the target scale is not described — we do not know what other scales were tested (4B? 8B? 10B?) or what the quality-vs-parameter Pareto curve looks like.

Mitigation Status: Not addressed. The paper presents the architectural choices as design decisions, not as hypotheses to be tested. The efficiency of the architecture relative to alternatives is asserted, not demonstrated through architectural ablation or scaling law analysis.

6.5 No Statistical Significance or Confidence Intervals on Benchmark Comparisons

The Assumption or Constraint: All benchmark comparisons in the paper are presented as point estimates without confidence intervals, standard deviations, or statistical significance tests. For benchmarks with small test sets — GenEval (likely a few hundred prompts, exact size not reported), DPG-Bench (1,065 prompts), WISE (1,000 prompts), and the custom CEdit-Bench (1,464 pairs) — the reported differences between top models are often small. For example:

  • GenEval (Table 2): LongCat-Image 0.87 vs. Qwen-Image 0.87 (tied) vs. Seedream 4.0 0.84 vs. HiDream-I1-Full 0.83
  • DPG-Bench (Table 3): LongCat-Image 86.80 vs. HunyuanImage-3.0 86.10 (0.70 point difference)
  • GEdit-Bench EN Overall (Table 10): LongCat-Image-Edit 7.64 vs. Qwen-Image-Edit 7.56 (0.08 point difference) vs. Seedream 4.0 7.68 (0.04 point difference)

The paper also does not report the number of human evaluators, inter-annotator agreement metrics, or confidence intervals for the MOS and SBS evaluations (Figures 14, 22).

The Consequence: For the closest comparisons — where the paper's claims of superiority or parity are most consequential — we cannot determine whether the reported differences are statistically reliable. At 1,000–1,500 test prompts, a difference of 0.5–1.0 percentage points or 0.08 GPT-4o-evaluated score units may be within the noise floor of sampling variability, prompt selection, or evaluation metric variance. Specifically:

  • The claim that LongCat-Image "matches Qwen-Image at 3.3× fewer parameters" on GenEval relies on an exact tie at 0.87. If the 95% confidence interval on GenEval scores is ±0.02 (plausible for a few hundred prompts), the tie could actually be a win or loss by a small margin.
  • The claim that LongCat-Image-Edit achieves "SOTA among open-source models" on GEdit-Bench relies on a 0.08 Overall score advantage over Qwen-Image-Edit. If the standard deviation of the GPT-4o evaluation is 0.1–0.2 score units, this difference is not statistically significant.
  • The human evaluation SBS results (Figure 22) show win rates of ~57% vs. ~43% — a 14 percentage point margin — but without knowing the sample size (400) and the annotator agreement, we cannot determine whether this reflects a reliable preference or could be reversed with different annotators or prompts.

What Evidence Exists in the Paper: None. No confidence intervals, error bars, standard deviations, or significance tests are reported for any quantitative result. The human evaluation methodology section (5.2) does not specify the number of annotators, annotation interface, or agreement metrics.

Mitigation Status: Not addressed. This is standard practice in many industry technical reports, where the emphasis is on demonstrating competitive performance rather than on rigorous statistical comparison. However, for a paper that makes precise claims about superiority (e.g., "outperforming Qwen-Image" on specific benchmarks), the absence of uncertainty quantification weakens those claims to the level of "suggestive evidence" rather than "demonstrated fact" when margins are small.

6.6 The Editing and T2I Models Are Separated Due to Unresolved Data Quality Conflicts

The Assumption or Constraint: The paper initially aimed to unify T2I and image editing into a single model but was forced to separate them due to a fundamental tension in training data quality. Section 6.4 states:

"Initially, we aim to unify T2I and image editing into a single model to leverage potential task synergies and minimize deployment costs. However, experiments reveal a critical data quality mismatch: the heavy reliance on synthetic data during editing pre-training noticeably degrades the photorealism of T2I generation compared to models trained solely on real data."

This is an acknowledged tradeoff that the paper does not resolve. The synthetic editing data (generated by expert models, video frame extraction, and interleaved corpus mining — Section 6.1) is necessary for teaching diverse editing capabilities but actively damages the photorealism that the T2I training pipeline so carefully cultivates through AIGC exclusion and quality curation. The consequence is the deployment of two separate models — LongCat-Image for T2I, LongCat-Image-Edit for editing — which doubles storage requirements, VRAM footprint for multi-task serving, and maintenance burden.

The Consequence: For a practitioner building a product that requires both generation and editing capabilities, the separation imposes:

  • 2× model storage and loading costs: Both 6B models must be stored and loaded, consuming ~12B parameters worth of VRAM if both need to be available simultaneously.
  • No knowledge sharing between tasks: The editing model cannot benefit from continued improvements to the T2I model, and vice versa, because they diverge after the mid-training checkpoint.
  • A fundamental uncertainty about the path to unification: The paper suggests that "by substituting synthetic datasets with large-scale interleaved corpora, future iterations can successfully merge these capabilities" (Section 6.4), but provides no evidence that such interleaved corpora exist at sufficient scale or quality to replace synthetic editing data. The unification roadmap is speculative.

This undermines a key part of the paper's efficiency narrative. One of the stated motivations is to reduce deployment costs compared to large models, but separating generation and editing into two 6B models — while individually efficient compared to 20B+ models — still requires more total parameters at serving time than a hypothetical unified 8–10B model that handled both tasks.

What Evidence Exists in the Paper: The paper is transparent about this limitation in Section 6.4, explicitly stating that the separation is "a data-driven issue, not an architectural flaw." The editing and T2I benchmarks (Tables 2–11) are reported for separate models, confirming the split.

Mitigation Status: Acknowledged and attributed to a solvable (but currently unsolved) data problem. The paper frames this as future work contingent on improved interleaved corpus mining, but provides no timeline, no evidence that the approach would work, and no analysis of how much synthetic editing data would need to be replaced to eliminate the photorealism degradation. For current deployment, the limitation is unresolved and users must accept the cost of serving two models.

7. Implications and Future Directions

How This Work Changes the Landscape

This paper shifts the conversation around image generation model development from parameter scaling as the default path to improvement toward data-centric engineering and targeted failure-mode intervention as viable alternatives. The conceptual shift is not that smaller models can be good — that was always theoretically possible — but that a 6B model can achieve competitive-or-superior performance to 20B–80B models on specific, commercially important capabilities (Chinese text rendering, photorealism, editing consistency) through interventions that parameter scaling alone cannot deliver.

The magnitude of this shift is a significant reframing, not a paradigm shift. The paper does not introduce a new architectural paradigm (the FLUX.1-dev backbone is adopted without fundamental modification) or a new training algorithm (DPO, GRPO, and MPO are existing methods). What it introduces is a diagnostic framework: identify the specific failure modes that limit current models (AIGC-contaminated textures, long-tail character coverage, visual consistency in editing), then design targeted interventions at the data, tokenization, and reward model levels to address each failure mode. This is a different philosophy from the scaling approach, which implicitly assumes that larger models trained on more data will resolve failure modes through emergent capability.

The paper resolves a specific tension in the literature: why do models keep getting larger without proportional improvements in photorealism and text rendering? The diagnostic answer — AIGC contamination in training data actively constrains models away from photorealistic outputs, and subword tokenization makes long-tail character rendering combinatorially harder — provides a mechanistic explanation for diminishing returns from scaling. Prior work had documented that scaling helps but had not isolated these specific bottlenecks. The paper's ChineseWord benchmark (Table 7) makes this starkly visible: Seedream 4.0, a commercial model trained on presumably massive compute, achieves 2.3% accuracy on Level 3 characters, while LongCat-Image at 6B achieves 70.3%. This 68-point gap cannot be explained by model capacity — it reflects fundamentally different approaches to the tokenization and data problems.

The research directions that become more attractive after this work:

  • Data forensics and quality curation over architectural innovation. If AIGC contamination is a primary photorealism bottleneck, then better AIGC detection, provenance tracking, and training data hygiene become first-class research problems rather than engineering afterthoughts.
  • Tokenization-level interventions for structured generation tasks. The character-level tokenization success suggests that similar granularity adjustments could benefit other generation tasks where the standard subword tokenizer fragments important conditioning signals — mathematical notation, chemical formulas, sheet music, code with specific formatting requirements.
  • Multi-dimensional reward modeling for RL. The ensemble reward approach (AIGC detection + OCR accuracy + distortion detection + human preference) demonstrates that targeted adversarial signals can address failure modes that general preference models miss. This opens a research direction in designing diagnostic reward ensembles for specific generation tasks.
  • Mid-training as a distinct model state worth optimizing and releasing. The paper's tripartite release strategy (mid-training Developer Version, final aligned model, editing model initialized from mid-training) suggests that intermediate training stages have independent value that the field has underexploited.

Research directions that become less attractive:

  • Brute-force parameter scaling for image generation without accompanying data quality interventions. The paper demonstrates that an 80B MoE model (HunyuanImage-3.0) can be substantially outperformed by a 6B model on multiple benchmarks (GenEval: 0.72 vs. 0.87; ChineseWord: 49.3% vs. 90.7%). Scaling without solving the data contamination and tokenization problems is empirically shown to yield diminishing returns.
  • Specialized text encoders for glyph rendering (e.g., GlyphByT5). The paper shows that a simple tokenizer modification to an existing LLM-based encoder achieves dramatically better Chinese character coverage than specialized encoder approaches, suggesting that the complexity of dedicated text rendering modules may be unnecessary if the base encoder is sufficiently powerful and the tokenization is appropriate.

Follow-Up Research This Work Enables

Ablation of isolated components to determine which interventions drive which improvements. The most important missing piece in this paper is the absence of controlled experiments that isolate the contribution of each design choice. A strong follow-up study would train a family of models that systematically ablate: (a) AIGC filtering from pre-training data (train with and without, measure photorealism via human evaluation); (b) AIGC detection as an RL reward model (train RL with and without this reward component, measure photorealism); (c) character-level vs. subword tokenization (train two models with identical data and architecture, differing only in tokenizer configuration for quoted text, measure ChineseWord accuracy); (d) multi-granularity captioning vs. single-granularity (train with only Photographic Level captions vs. the weighted 4-level mixture, measure DPG-Bench and human evaluation alignment); (e) mid-training vs. post-training initialization for editing (train two editing models from different checkpoints with identical editing data, measure CEdit-Bench and GEdit-Bench). This ablation suite would transform the paper's contributions from "this combination works" to "these specific components matter this much for these specific outcomes," which is what the field needs to adopt or adapt the techniques.

Does the AIGC contamination hypothesis generalize across model architectures and training scales? The paper's central diagnostic — that even small amounts of AIGC data in pre-training cause photorealism collapse — is asserted based on internal observations but never demonstrated in a controlled experiment. A strong follow-up would train diffusion models at multiple scales (1B, 3B, 6B, 12B) on identical data with varying levels of AIGC contamination (0%, 1%, 5%, 10%, 25%) and measure photorealism via human evaluation MOS and automated metrics (FID against a real-photo reference set, perhaps). Key measurements: at what contamination level does photorealism degradation become statistically detectable? Does the degradation threshold depend on model scale (are larger models more or less susceptible)? Can the degradation be reversed by subsequent training on clean data, or is it path-dependent (once the model collapses to the synthetic manifold, it cannot escape)? This experiment would validate or refute one of the paper's most consequential claims and provide practical guidance for data curation in future training pipelines.

Can the character-level tokenization approach be extended to other structured-generation tasks? The paper demonstrates that switching from subword to character tokenization for quoted text spans dramatically improves Chinese character rendering (ChineseWord Level 3: 70.3% vs. 2.3–6.1% for competitors). This suggests a broader hypothesis: when a generation task requires precise rendering of structured symbolic content, tokenization granularity matching the structure's atomic units eliminates a combinatorial bottleneck. Strong follow-up studies would test this on: (a) mathematical formula rendering (LaTeX in images), comparing standard LLM tokenization against symbol-level tokenization for math expressions; (b) code snippet rendering in images (indentation, keyword coloring), comparing subword tokenization against character-level or token-level tokenization of code blocks; (c) sheet music or musical notation rendering, comparing standard tokenization against note-level or symbol-level encoding; (d) multi-language text rendering (Japanese kanji/kana, Korean hangul/hanja, Arabic script), testing whether character-level tokenization for the target script improves rendering accuracy for each language's long-tail characters. For each, the experimental design would mirror the paper's approach: train models with and without the tokenization change, keeping data and architecture identical, and measure accuracy on a systematically constructed benchmark covering common and rare symbols.

Can interleaved web corpora replace synthetic editing data to enable a unified generation + editing model? The paper was forced to separate its T2I and editing models due to synthetic editing data degrading photorealism (Section 6.4). The authors hypothesize that "by substituting synthetic datasets with large-scale interleaved corpora, future iterations can successfully merge these capabilities into a unified model without sacrificing generation quality." Testing this hypothesis requires: (a) mining a large-scale interleaved corpus (image-text-image sequences from the web) where consecutive images have natural editing relationships (same scene, different time; same product, different angle; before/after transformations); (b) training a unified model on this corpus plus T2I data, compared against the separate-model approach, measuring both editing quality (CEdit-Bench, GEdit-Bench) and photorealism (human evaluation MOS on T2I outputs); (c) determining the minimum ratio of real interleaved data to synthetic editing data needed to avoid photorealism degradation. Success would eliminate the deployment cost of serving two separate models and would validate the interleaved corpus mining direction that the paper identifies as a "critical direction for long-term data engineering" (Section 6.1.4).

Adversarial AIGC detection as a general-purpose photorealism reward: does it transfer across domains and model families? The paper's use of an AIGC detector as an RL reward model — adversarially pushing the generator away from synthetic-looking outputs — is conceptually novel. A strong follow-up would test whether this technique generalizes: (a) train an AIGC detector on a diverse set of AI-generated images from multiple model families (SDXL, FLUX, Midjourney, DALL-E), then use it as a reward model when training a different model family (e.g., use it to RL-tune a U-Net-based diffusion model); (b) measure whether the photorealism improvement transfers or whether the detector overfits to artifacts specific to the models it was trained on; (c) test in non-photorealistic domains (illustration, 3D rendering, medical imaging) where "realism" means different things — does an AIGC detector trained on photorealistic images penalize stylistic outputs that are intentionally non-photorealistic? This experiment would determine whether AIGC-adversarial training is a general technique or a domain-specific hack.

Dynamic difficulty-adaptive synthetic data sampling for long-tail generation tasks. The paper's dynamic sampling strategy for synthetic text data (Section 4.1) — increasing sampling probability for characters with high error rates, phasing out synthetic data once characters are learned — is a conceptually elegant approach to curriculum learning for long-tail distributions. A strong follow-up would formalize and generalize this: (a) define a general framework where a "difficulty estimator" (e.g., per-class validation accuracy) controls the mixing ratio between synthetic focused-training data and real diverse data; (b) test on other long-tail generation problems (rare object categories in image generation, uncommon programming languages in code generation, infrequent scientific entities in diagram generation); (c) compare against static mixing ratios and against uniform synthetic data strategies to measure the benefit of adaptivity; (d) determine whether the phase-out schedule (linear decay, step function, accuracy-thresholded) matters for final performance. The paper provides anecdotal evidence that the dynamic strategy is beneficial, but no controlled comparison against static sampling, leaving the technique's generalizability unvalidated.

Practical Applications and Downstream Use Cases

Chinese-language poster and advertisement generation. LongCat-Image's ChineseWord benchmark result (90.7% overall accuracy, 70.3% on rare Level 3 characters) combined with its Poster benchmark performance (92.0% accuracy, Table 8) directly enables automated generation of Chinese-language marketing materials that include accurate rendering of brand names, product specifications, legal disclaimers, and decorative text — all of which may contain rare characters. For a Chinese e-commerce platform generating product listing images, the difference between 90.7% character accuracy and the next-best open-source model's 56.6% (Qwen-Image) translates to roughly one in three images having text rendering errors versus fewer than one in ten. At the scale of millions of product listings, this reliability gap determines whether the generation pipeline can be automated or requires manual verification. The 6B parameter count means this can run on consumer GPUs (likely 16–24GB VRAM), enabling deployment at individual merchant scale rather than requiring centralized datacenter infrastructure.

On-device or edge-deployed image editing for photography applications. The editing model's performance on fine-grained portrait editing (Figure 24: blemish removal, hairstyle modification, lighting adjustment, ID photo generation) and human-centric editing (Figure 25: pose modification, viewpoint transformation, lighting simulation) at 6B parameters makes it feasible to deploy high-quality photo editing on devices without cloud round-trips. The benchmark results show that LongCat-Image-Edit achieves SOTA among open-source models on CEdit-Bench (7.67 Overall EN score vs. Qwen-Image-Edit's 7.52, Table 9) and ImgEdit-Bench (4.50 Overall vs. 4.27, Table 11), with a 56.3% consistency win rate against Qwen-Image-Edit (Figure 22). For a mobile photography application processing user photos, the 6B parameter count means the model could run on-device with quantization (int8 or int4), eliminating the latency and privacy concerns of cloud-based editing. The joint training with T2I data (Section 6.3.1) suggests the editing model retains some generative capabilities, potentially enabling use cases like "remove this object and fill the background naturally" that require both editing precision and generative inpainting.

Multilingual content creation pipelines where rare character coverage is critical. The ChineseWord benchmark's coverage of 8,105 characters spanning the General Standard Chinese Characters Table makes LongCat-Image suitable for applications that cannot tolerate character rendering failures: textbook illustration generation, legal document visualization, cultural heritage preservation (rendering historical texts with archaic characters), and multilingual signage design. The 70.3% accuracy on Level 3 characters (the rarest tier) means the model correctly renders the large majority of even the most obscure characters in the standard Chinese lexicon. For a textbook publisher generating illustrated vocabulary cards for rare characters, the alternative — Seedream 4.0 at 2.3% Level 3 accuracy — would produce incorrect renderings for 98% of rare characters, requiring manual correction of nearly every image. The efficiency of the 6B architecture means this capability is accessible to organizations without massive compute budgets, which is particularly relevant for educational and cultural institutions.

Base model for downstream fine-tuning via the Developer Version release. The mid-training checkpoint release (Section 4.2) as a "Developer Version" that retains "high plasticity and adaptability" makes LongCat-Image attractive as a starting point for domain-specific fine-tuning. Organizations that need specialized generation capabilities — medical image generation, architectural visualization, fashion design, food photography — can start from a checkpoint that has strong photorealistic foundations (from the curated mid-training data) and broad semantic knowledge (from pre-training) but is not locked into a specific aesthetic (no SFT/RL alignment). The paper demonstrates this paradigm with the editing model (initialized from mid-training, achieving SOTA open-source editing performance), but the same approach could apply to any specialized generation domain. The availability of the full training codebase further lowers the barrier to adapting the training pipeline for domain-specific data.

When to Prefer This Method

The paper positions LongCat-Image against specific named alternatives — Qwen-Image (20B), HunyuanImage-3.0 (80B MoE), Seedream 4.0 (commercial), and various open-source editors — and the performance boundaries are clear enough to support explicit decision rules.

Prefer LongCat-Image (or its approach) when:

  • Chinese text rendering with broad character coverage is a primary requirement. The ChineseWord benchmark (Table 7) shows LongCat-Image achieving 90.7% overall vs. 56.6% for the next-best open-source model and 58.5% for the commercial Seedream 4.0. On rare Level 3 characters, the gap is 70.3% vs. 2.3–6.1%. If your application involves rendering Chinese names, locations, technical terms, or classical texts that may contain infrequent characters, LongCat-Image is the best available option by a wide margin — open-source or commercial.

  • Deployment on consumer-grade or edge hardware is required. At 6B parameters, LongCat-Image is 3.3× smaller than Qwen-Image (20B) and 13× smaller than HunyuanImage-3.0 (80B MoE), with correspondingly lower VRAM requirements. If your deployment target is a single consumer GPU (16–24GB VRAM), a 20B+ model may not fit, making LongCat-Image the strongest option that can actually run. The paper does not provide VRAM benchmarks, but a 6B diffusion transformer with a 7B text encoder can be expected to run in 16–24GB at FP16 and potentially under 12GB with int8 quantization.

  • Open-source image editing with strong visual consistency is needed. LongCat-Image-Edit achieves SOTA among open-source models on CEdit-Bench, GEdit-Bench, and ImgEdit-Bench (Tables 9–11), with human evaluation win rates of ~57% against the strongest open-source alternative (Qwen-Image-Edit) on both comprehensive quality and consistency (Figure 22). If you need editing capabilities and cannot use commercial APIs (due to cost, privacy, or customization requirements), LongCat-Image-Edit is the best currently available open-source option.

  • You need a base model for downstream fine-tuning that balances quality with plasticity. The mid-training Developer Version (Section 4.2) is specifically designed for this use case — it has better quality than a raw pre-trained model (due to curated mid-training data) but broader generative capabilities than a fully aligned model (no SFT/RL mode collapse). If you are fine-tuning for a specialized domain (medical, architectural, fashion), starting from the mid-training checkpoint is likely preferable to starting from a pre-trained model (lower initial quality) or an aligned model (less adaptable).

Prefer larger models or commercial alternatives when:

  • Complex multi-text-region rendering is the primary task. The CVTG-2K benchmark (Table 6) shows LongCat-Image's word accuracy degrading from 0.9129 at 2 regions to 0.8310 at 5 regions, while Seedream 4.0 maintains 0.8898 → 0.8873. For applications requiring multiple distinct text elements in different image regions (complex posters, infographics), the larger model maintains consistency better as text count increases.

  • The broadest semantic coverage on diverse prompts is required. On DPG-Bench (Table 3), LongCat-Image trails Qwen-Image by 1.52 points overall (86.80 vs. 88.32), with the largest gap in the "Other" category (87.50 vs. 92.73) representing miscellaneous complex semantic relationships. If your prompt distribution is highly diverse and you need the best average performance across all semantic categories, the 20B model maintains an edge.

  • You need a unified generation + editing model in a single deployment. LongCat-Image separates T2I and editing into two 6B models due to unresolved data quality conflicts (Section 6.4), requiring 12B total parameters if both capabilities are needed simultaneously. A unified 20B model that handles both tasks (if available) may actually have lower total deployment cost than two separate 6B models, and the paper provides no timeline for when unified capabilities will be achieved.

  • You are competing with commercial editing systems and every percentage point matters. The SBS human evaluation (Figure 22) shows LongCat-Image-Edit winning only ~43% of comparisons against Seedream 4.0 and ~43% against Nano Banana on comprehensive quality. If your application requires editing quality at or above the commercial state-of-the-art, open-source models (including LongCat-Image-Edit) have not yet closed that gap.