HiAR: Efficient Autoregressive Long Video Generation via Hierarchical Denoising

ArXiv: 2603.08703

🎯 Pitch

Conditioning each video block on previous blocks at the same noise level instead of fully clean output prevents the runaway error accumulation that normally degrades autoregressive long video. When paired with a forward-KL regularization to preserve motion diversity, this matched-noise conditioning achieves the lowest temporal drift reported on VBench (0.257) while also enabling a 1.8× inference speedup through pipelined parallelism.

---

1. Executive Summary

This paper introduces HiAR, a hierarchical denoising framework for autoregressive long video generation that reverses the conventional block-first pipeline—instead of fully denoising each video block before generating the next, HiAR performs causal generation across all blocks within each denoising step, conditioning each block on context at a matched noise level rather than on predicted clean frames. Built on the Wan2.1-1.3B backbone and evaluated on VBench with a 4-step denoising schedule, the method combines hierarchical denoising with a forward-KL regularizer in bidirectional-attention mode to counteract the low-motion shortcut inherent to reverse-KL distillation. HiAR achieves the best overall VBench score (0.821), the lowest temporal drift (0.257, a 27.6% reduction over Self-Forcing), and a ~1.8× wall-clock speedup via pipelined parallelism, establishing that matched-noise context provides sufficient signal for temporal continuity while attenuating error propagation—but only when paired with train–test alignment under the hierarchical pipeline and explicit motion-diversity regularization, since hierarchical denoising applied at inference alone breaks continuity and forward-KL omission causes dynamics to collapse to 0.445.

2. Context and Motivation

The Core Problem: Scaling Video Generation to Arbitrary Lengths While Maintaining Quality

The fundamental challenge this paper tackles is straightforward to state but has proven remarkably difficult to solve: how do we generate videos of theoretically unlimited length without watching their quality degrade over time? This matters because current state-of-the-art video diffusion models, despite producing impressive short clips lasting a few seconds, hit a hard ceiling imposed by their underlying architecture. Understanding why this ceiling exists—and why the obvious solution creates its own problems—requires examining the landscape of video generation approaches and the specific failure modes that emerge when we try to push beyond fixed-duration outputs.

The significance of this problem extends well beyond the convenience of longer videos. The paper situates long-horizon video generation as "central to interactive agents and world models," citing recent work on interactive video environments (He et al., 2025; Ye et al., 2025; Mao et al., 2025; Sun et al., 2025; Hong et al., 2025; Tang et al., 2026). These applications—real-time interactive worlds, persistent visual environments for embodied agents, streaming video interfaces—demand continuous, indefinitely sustained video output that maintains physical and visual consistency. A world model that slowly drifts into oversaturation or motion collapse over tens of seconds cannot serve as a reliable substrate for long-horizon planning or interaction. The problem is therefore not merely aesthetic; it is a foundational requirement for a class of emerging AI systems.

The Bidirectional Bottleneck

To understand why this is hard, we first need to understand the dominant paradigm in modern video generation. Bidirectional-attention diffusion models—examples include Sora (OpenAI, 2025), Wan2.1 (Team, 2025), Kling (Kling, 2025), Veo (Google, 2025), and Gen-4.5 (Runway, 2025)—operate by jointly denoising all frames of a video simultaneously. Every frame attends to every other frame during the denoising process. This global attention mechanism is remarkably effective at maintaining temporal coherence: frames remain consistent with one another because the model can look forward and backward across the entire sequence at once.

However, this strength is also the architecture's fundamental limitation. Joint attention across all frames means the computational cost scales quadratically with sequence length, and more importantly, the model was trained on a fixed temporal window. You cannot simply extend a bidirectional model to generate a 5-minute video if it was only ever trained on 5-second clips. The paper presents this concisely: bidirectional models "cannot easily scale to arbitrary durations" (Section 2.2). They produce videos of fixed, predetermined length with no mechanism for extension beyond their training window. This is the "fixed horizon" problem illustrated in Figure 1(a): Wan2.1 generates a high-quality 4-second clip with good continuity, but that is all it can do—the generation ends there.

The Autoregressive Solution and Its Hidden Pathology

The natural escape from the fixed-horizon bottleneck is to generate video autoregressively—break the desired long video into a sequence of shorter blocks, generate the first block, then generate the second block conditioned on the first, then the third conditioned on the first two, and so on. This paradigm, called causal autoregressive (AR) generation, offers everything bidirectional models cannot: streaming output (frames appear sequentially rather than all at once), indefinite extension (just keep generating more blocks), real-time interaction (intervene at block boundaries), and theoretically infinite length. These properties make AR generation attractive for the interactive agent and world model applications the paper cites (Section 2.2).

The challenge, and the central problem this paper addresses, is that AR generation introduces a distribution drift problem—a progressive degradation in video quality that compounds with each successive block. The paper's Figure 1(b) illustrates this vividly: over a 20-second generation using standard AR methods (what the paper calls "Self-Forcing"), visual quality degrades visibly, with the later segments showing the telltale symptoms described in the introduction: "oversaturation, over-sharpening, motion repetition, and semantic drift" (Section 1).

To understand why this happens, we need to examine precisely how conventional AR video generation handles context—the previously generated frames that condition the next block's generation.

The Context Noise Level Problem

When generating block $B_n$ (the $n$-th segment of video), the model conditions its denoising process on the previously generated blocks $B_1, \ldots, B_{n-1}$. The key question is: at what noise level should this context be provided? The existing AR methods, including Self-Forcing (Anonymous, 2025), provide the context at noise level $t_c = 0$—meaning the previously generated frames are fully denoised, perfectly clean predictions of what those blocks should look like.

At first glance, this seems entirely reasonable. Clean frames contain maximum information about what came before, providing the strongest possible conditioning signal for temporal consistency. If you want the next block to flow naturally from the previous ones, shouldn't you give the model the clearest possible picture of those previous frames?

The paper's central insight—the one that motivates the entire hierarchical denoising architecture—is that this reasoning contains a hidden flaw: clean context propagates errors with maximum fidelity. When block $B_{n-1}$ is generated, it contains some prediction error $\delta^{(n-1)}$ relative to what a perfect model would produce. When this block is passed to the denoiser for block $B_n$ at noise level $t_c = 0$, the context is simply the clean (but imperfect) prediction $\hat{x}^{(n-1)}_0$. Every error in the previous block is presented to the model at full strength, with no attenuation. The model then conditions on these errors as if they were ground truth, and produces block $B_n$ accordingly—its own errors compounding with the inherited ones. This process repeats for block $B_{n+1}$, and so on, with errors accumulating along the chain.

The paper formalizes this in Section 3.1 through a bias-information trade-off that we will examine in detail later, but the core intuition is accessible without mathematics: noisy context carries less information about the previous frames, which sounds bad, but it also carries less information about the errors in those frames. The errors are diluted by noise, reducing their propagation to subsequent blocks.

Prior Approaches and Where They Fall Short

The paper identifies a progression of attempts to solve the AR drift problem, each addressing part of the issue but leaving a critical gap.

Teacher Forcing and the Train-Test Mismatch

The simplest approach to training an AR generator is teacher forcing (Williams and Zipser, 1989): during training, condition each block on the ground-truth clean previous frames from the training data, not on the model's own generated outputs. This works beautifully during training because the model always sees perfect context—there are no errors to propagate. At inference time, however, the model must condition on its own (imperfect) generated frames, which it has never practiced doing. This train–test mismatch, known as exposure bias (Bengio et al., 2015), means the model is unprepared for the error-accumulation dynamics it encounters during actual generation. Small initial errors cascade into large deviations because the model has never learned to recover from or account for imperfect context.

The paper cites multiple works that have documented this failure mode: Gao et al. (2024), Hu et al. (2024), Jin et al. (2024a), Zhang et al. (2025) all observe that teacher-forced training produces models that degrade under autoregressive inference, manifesting as the progressive oversaturation, motion repetition, and semantic drift that Figure 1(b) illustrates.

Diffusion Forcing: Training with Noisy Contexts

Diffusion Forcing (Chen et al., 2024; Yin et al., 2025b; Chen et al., 2025b; Gu et al., 2025; Teng et al., 2025b; Song et al., 2025; Po et al., 2025) partially addresses the train–test gap by training with independent per-token noise levels. Rather than always training with clean context, the model is exposed to contexts at various noise levels during training, so it learns to denoise under heterogeneous noise conditions. This gives the model some robustness: at inference time, when it encounters imperfect (partially noisy) context from its own previous predictions, it has at least some training experience with non-clean conditioning signals.

Diffusion Forcing represents a significant conceptual advance—it recognizes that the train–test gap is fundamentally about the context noise distribution—but it does not solve the fundamental problem of error propagation. During training, the noisy contexts are still derived from ground-truth data (just with added independent noise), not from the model's own prediction errors. The model learns to handle noisy-but-unbiased context, not context that contains correlated, compounding prediction errors of the kind that actually arise during autoregressive rollouts. The distribution of errors the model encounters during inference remains out-of-distribution relative to training.

Self-Forcing: Closing the Gap with Self-Rollout Training

Self-Forcing (Anonymous, 2025; Yin et al., 2024a,c; Yi et al., 2025) takes the next logical step: close the train–test gap entirely by training on the model's own rolled-out predictions. During each training iteration, the model first generates block $B_{n-1}$ using its current parameters (the "student" model), then uses that generated block as context for training the denoising of block $B_n$. The context the model sees during training is therefore drawn from exactly the same distribution it will encounter during inference—its own (imperfect) predictions.

This is a substantial improvement. Self-Forcing trains the model to recover from and account for its own characteristic error patterns, and the paper acknowledges that it "achieves notable improvements at moderate horizons" (Section 2.2). However, Self-Forcing still makes one critical design choice that the paper identifies as the root cause of persistent degradation: it provides the rolled-out context at noise level $t_c = 0$—perfectly clean, full-fidelity predictions. Even though the model now trains on its own errors, those errors are still presented as if they were ground truth, with no noise attenuation. The model learns to cope with its own error patterns, but it still propagates them forward at full strength. The paper's Figure 1(b) shows that Self-Forcing, while better than naive teacher forcing, still exhibits visible degradation over 20 seconds—oversaturation, hue shifts, and loss of detail accumulate noticeably by the end of the sequence.

The training objective for Self-Forcing is an asymmetric Distribution Matching Distillation (DMD) loss (Yin et al., 2024b,d), formulated as a reverse KL divergence between the student's one-step output distribution and a teacher model's multi-step output distribution. This objective, which we will examine in Section 3, encourages the student to match the teacher's output distribution. The paper identifies this reverse-KL objective as a source of additional problems—specifically, a tendency toward mode collapse where the model learns to generate low-motion, near-static videos because these are easier to denoise and less prone to compounding errors during rollout. We will return to this "low-motion shortcut" when examining the training methodology.

The Gap This Paper Addresses

The landscape prior to this work can be summarized as follows: autoregressive video generation is the only viable path to indefinite-length video, but it suffers from a progressive quality degradation driven by error accumulation. Previous solutions have addressed pieces of the puzzle—Diffusion Forcing improved robustness to noisy contexts, Self-Forcing closed the train–test distribution gap—but all of them inherit the same fundamental assumption: context should be provided at as clean a noise level as possible, typically $t_c = 0$. None of the prior approaches questioned whether clean context is actually optimal.

The paper's core challenge to this assumption comes from an empirical observation about bidirectional diffusion models. As illustrated in Figure 1(a), bidirectional models like Wan2.1 denoise all frames concurrently from a shared, high noise level at the start of generation, yet still produce temporally coherent videos. This demonstrates that noisy context already provides sufficient signal for temporal continuity. The bidirectional model never has access to clean frames during its early denoising steps—every frame is equally noisy—yet the final output is coherent because the model can infer temporal relationships from the shared noise structure.

This observation motivates the paper's key proposition: if noisy context is sufficient for temporal continuity in the bidirectional setting, perhaps it is not only sufficient but actually preferable in the autoregressive setting, because noisy context attenuates error propagation. A block conditioned on noisy previous frames inherits less of the previous blocks' prediction errors, because those errors are diluted by the noise—they share the same attenuation coefficient $(1 - \sigma_{t_c})$ as the true signal, as formalized in Equation 7.

How This Paper Positions Itself

The paper's contribution is most accurately understood not as an entirely new method for video generation, but as a fundamental rethinking of the generation order in autoregressive pipelines, motivated by an analytic insight about optimal context noise levels, and supported by two complementary innovations that make the reordering practically viable.

The analytic contribution (Section 3.1) is a formal error decomposition showing that the context noise level $t_c$ controls a bias–information trade-off: higher $t_c$ reduces propagated bias but also reduces useful signal; lower $t_c$ preserves signal but amplifies error propagation. The paper derives that the optimal noise level is $t_c^* = t_{j+1}$—the output noise level of the current denoising step—which is the noisiest context that still satisfies temporal causality (the constraint that context must be at least as informative as the current block's state after the denoising step). This provides a principled justification for departing from the $t_c = 0$ convention.

The architectural contribution is Hierarchical Denoising (Section 3.2): instead of the conventional block-first order (fully denoise block 1, then fully denoise block 2, etc.), HiAR performs a step-first order (for denoising step 1, generate all blocks; for step 2, generate all blocks; etc.). At each step $j$, block $B_n$ conditions on block $B_{n-1}$ at noise level $t_{j+1}$—the matched-noise level identified as optimal. This reordering is what the paper calls "simple yet fundamental" (Section 1): it is conceptually straightforward, but it inverts the entire generation pipeline and has consequences for both quality and efficiency.

Critically, the paper demonstrates that this reordering is not something that can simply be applied at inference time to an existing model. Figure 1(c) shows that applying hierarchical denoising at inference only ("w/o training") does reduce drift compared to Self-Forcing, but at a severe cost to visual quality and temporal continuity. The model was trained under the block-first paradigm, where it always saw clean or nearly-clean context; asking it to operate with matched-noise context at test time creates a new train–test mismatch. The paper therefore retrains the model from scratch under the hierarchical denoising schedule, ensuring the model learns to generate under exactly the conditions it will encounter during inference.

The regularization contribution (Section 3.3) emerges from the observation that retraining under hierarchical denoising amplifies a low-motion shortcut: the mode-seeking reverse-KL objective (DMD) encourages the model to produce outputs that are easy to denoise and unlikely to cause compounding errors—namely, near-static videos with minimal motion. Hierarchical denoising makes the learning problem harder (the model must condition on contexts at varying noise levels rather than always-clean ones), requiring more training steps and giving the mode-seeking objective more iterations to collapse onto the low-motion mode. The paper introduces a forward-KL regularizer computed in bidirectional-attention mode to counteract this collapse. The key design insight is that motion dynamics under bidirectional and causal attention are strongly positively correlated (Pearson $r = 0.968$, Figure 4), so regularizing the bidirectional mode effectively constrains causal-mode dynamics without interfering with the DMD loss that operates in causal mode.

Connections to Broader Research Themes

The paper positions HiAR within several interconnected research threads that help contextualize its significance:

The bidirectional-to-AR distillation trend. Several recent works (CausVid from Yin et al., 2025a; Self-Forcing from Anonymous, 2025; Causal Forcing from Zhu et al., 2025) share a common approach: take a high-quality bidirectional diffusion model (Wan2.1), distill it into a fast, few-step autoregressive generator. This paradigm is attractive because it inherits the strong pretrained representations of the bidirectional teacher while gaining the streaming, indefinite-extension capabilities of AR generation. HiAR builds directly on this lineage—it uses the same Wan2.1-1.3B backbone, the same 4-step denoising schedule, and the same DMD distillation objective as Self-Forcing—but changes the generation order and adds the forward-KL regularizer. This makes the comparisons in Table 1 particularly meaningful: HiAR is evaluated against methods that share its foundation model and compute budget, isolating the effect of the hierarchical denoising architecture.

Efficiency through parallelism. The hierarchical structure creates an opportunity that the block-first paradigm cannot exploit: different blocks at the same denoising step are independent of each other (block $B_n$ at step $j$ depends only on block $B_{n-1}$ at step $j$, which is already computed). The paper exploits this through pipelined parallelism, where denoising steps are assigned to dedicated processes and blocks flow through the pipeline, achieving a $\sim 1.8\times$ wall-clock speedup. This efficiency gain is not a separate contribution but a direct consequence of the hierarchical architecture, and it matters practically: the other distilled AR models in Table 1 all achieve 17 fps and 0.69 s latency, while HiAR reaches 30 fps and 0.30 s latency with the same model backbone and denoising steps.

The mode-seeking problem in distillation. The paper contributes to a broader understanding of distribution matching objectives in diffusion distillation. The observation that reverse-KL ($D_{KL}(p_\theta \| p_{\text{teacher}})$) encourages mode-seeking behavior—concentrating probability mass on a single high-density region of the teacher's distribution—is well-established in the generative modeling literature. The paper's contribution is showing that this manifests in video generation specifically as a low-motion collapse, where the model learns that static or near-static videos minimize the distillation loss because they are easier to denoise consistently across autoregressive rollouts. The forward-KL regularizer, computed via trajectory matching on teacher-generated samples, provides a principled counterbalance: forward-KL ($D_{KL}(p_{\text{teacher}} \| p_\theta)$) encourages mode-covering, penalizing the model for ignoring regions of the teacher's output distribution. The bidirectional-attention decoupling strategy—computing the forward-KL term in bidirectional mode while keeping the DMD loss in causal mode—is an elegant solution to what would otherwise be conflicting gradient signals.

Summary of the Motivation

The paper addresses a specific, well-defined gap: autoregressive video generation is necessary for indefinite-length video but suffers from error-propagation-driven quality degradation, and existing solutions fail because they condition on fully clean context, which propagates errors with maximum fidelity. The insight that motivates the solution—that noisy context attenuates error propagation while still providing sufficient information for temporal continuity—comes from observing that bidirectional models achieve coherence without clean-context access. Hierarchical denoising implements this insight by reversing the generation order, and the forward-KL regularizer makes the resulting training procedure stable by preventing motion collapse. The paper positions itself as extending the bidirectional-to-AR distillation paradigm while fundamentally rethinking how context should be provided during autoregressive rollouts.

3. Technical Approach

3.1 Reader Orientation (Approachable Technical Breakdown)

HiAR is a video generation system that produces arbitrarily long videos by generating short segments (blocks) one after another, but with a crucial twist on the standard autoregressive pipeline: instead of completing each block fully before starting the next, it performs one denoising step on all blocks, then the next denoising step on all blocks, and so on—like an assembly line where every station works on every part simultaneously rather than one station finishing a part completely before passing it along. The system solves the problem of progressive quality degradation in long video generation—the tendency for videos to become oversaturated, blurry, or semantically incoherent over time—by ensuring that when a new block conditions on previously generated blocks, those previous blocks are provided at a matched noise level rather than in fully cleaned form, which dilutes prediction errors and prevents them from compounding destructively across the autoregressive chain.

3.2 Big-Picture Architecture (Diagram in Words)

The HiAR system comprises five major components that interact during both training and inference:

1. Base diffusion model (Wan2.1-1.3B). A pretrained bidirectional-attention video diffusion transformer that serves as both the initialization for the student generator and the source of teacher signals for distillation. It operates on latent representations of video frames.

2. Hierarchical denoising scheduler. The inference procedure that determines the order of operations. Given $N$ blocks of video to generate and $S$ denoising steps, it processes an $N \times S$ grid along anti-diagonals: step 1 denoises all $N$ blocks from noise level $t_1$ to $t_2$, step 2 denoises all $N$ blocks from $t_2$ to $t_3$, and so on. At each step $j$, block $n$ conditions on block $n-1$ at the same output noise level $t_{j+1}$ via causal attention.

3. Causal self-rollout training loop. During training, the model (student) first generates block $n-1$ using its own current parameters, then uses that generated block as context to train the denoising of block $n$. This closes the train–test gap because the context distribution at training time matches the context distribution at inference time.

4. Distribution Matching Distillation (DMD) loss (reverse-KL). The primary training objective that encourages the student's single-step denoising output to match the teacher's multi-step output distribution. This is a mode-seeking objective that can cause motion collapse if used alone.

5. Forward-KL regularizer (bidirectional-attention trajectory matching). An auxiliary loss computed by running the teacher through a long denoising trajectory, extracting checkpoints at the student's schedule, and supervising the student to match each consecutive pair. This encourages mode-covering behavior and preserves motion diversity. Crucially, it is computed in bidirectional-attention mode only, leaving the causal DMD loss unmodified.

Information flows through the system as follows during inference: (1) all $N$ blocks are initialized with Gaussian noise at $t_1$; (2) for step $j = 1$, block 1 is denoised to $t_2$, then block 2 is denoised to $t_2$ conditioned on block 1 at $t_2$, then block 3 conditioned on block 2 at $t_2$, etc.; (3) the KV cache is updated with all blocks at $t_2$; (4) step $j = 2$ proceeds similarly, conditioning on context at $t_3$; (5) after $S$ steps, all blocks are at $t_S \approx 0$ and the generated video is complete.

During training, the same hierarchical schedule is used, but with an additional teacher model providing the DMD critic signal and the forward-KL trajectory targets, as illustrated in Figure 2.

3.3 Roadmap for the Deep Dive

The technical approach builds logically from an analytic insight to an architectural change to a training stabilization mechanism. We will proceed in this order:

• First, the error decomposition and optimal context noise level derivation (Section 3.1 of the paper)—this is the theoretical foundation that justifies everything that follows. Without understanding why $t_c = t_{j+1}$ is optimal, the hierarchical denoising procedure appears arbitrary.

• Second, the hierarchical denoising inference procedure (Section 3.2)—the concrete algorithm that implements the matched-noise conditioning principle, including the pipelined parallelism optimization that yields the speedup.

• Third, the training methodology (Section 3.3)—how the model is trained under the hierarchical schedule, why the reverse-KL DMD objective causes motion collapse, and how the forward-KL regularizer counteracts this with specific design choices (bidirectional-attention decoupling, early-step restriction, trajectory matching formulation).

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily an architectural innovation paper whose core idea is that the generation order in autoregressive video diffusion should be inverted from block-first to step-first, with each block conditioning on context at the matched output noise level rather than on fully denoised predictions, and that this reordering must be paired with retraining under the hierarchical schedule and motion-diversity regularization to be effective.

---

Error Decomposition and Optimal Context Noise Level

Framing the problem mathematically. The paper's key analytic contribution is a formal decomposition of the error propagation mechanism in autoregressive diffusion. Consider the generation of block $B_n$ at denoising step $j$, which moves the block from noise level $t_j$ to $t_{j+1}$ (where $t_1 > t_2 > \dots > t_S \approx 0$, so $t_j > t_{j+1}$). The model conditions on the previously generated block $B_{n-1}$, but does so by presenting that block's latent at some context noise level $t_c$ rather than at its clean state. The question is: what value of $t_c$ minimizes the error propagated to $B_n$ while still providing sufficient information for temporal coherence?

To answer this, the paper first defines the context construction. Let $x^{(n-1)}_0$ be the ground-truth clean latent of block $n-1$ (unknown at inference time) and $\hat{x}^{(n-1)}_0 = x^{(n-1)}_0 + \delta^{(n-1)}$ be the model's prediction, where $\delta^{(n-1)}$ represents the accumulated prediction error from all previous blocks and denoising steps. The context at noise level $t_c$ is formed by adding noise to the prediction:

$c^{(t_c)}_{n-1} = (1 - \sigma_{t_c}) \hat{x}^{(n-1)}_0 + \sigma_{t_c} \eta, \quad \eta \sim \mathcal{N}(0, I)$

where $\sigma_{t_c} = \frac{s \cdot t_c}{1 + (s - 1) \cdot t_c}$ is the noise schedule from Equation 1, $s > 0$ is a shift parameter controlling schedule curvature, $t_c \in [0, 1]$ is the continuous time, $\eta$ is independent standard Gaussian noise, and $\hat{x}^{(n-1)}_0$ is the model's (imperfect) prediction of the clean previous block.

What it computes: this equation constructs the actual tensor that will be fed to the model as context for block $n-1$ when denoising block $n$. It interpolates between the predicted clean frame $\hat{x}^{(n-1)}_0$ and random noise $\eta$, with the interpolation coefficient $\sigma_{t_c}$ determining how much noise is added. When $t_c = 0$, $\sigma_0 = 0$, so the context is purely the clean prediction with no noise. When $t_c = 1$, $\sigma_1 \approx 1$, so the context is almost pure noise with minimal signal from the prediction.

Why this form: this is the standard forward noising process from the diffusion/flow matching framework (Equation 1), applied to the predicted clean frame rather than the ground-truth clean frame. Using the same noise schedule as the denoising process ensures consistency—the context noise levels are drawn from the same distribution the model was trained to reverse.

Error decomposition. Substituting $\hat{x}^{(n-1)}_0 = x^{(n-1)}_0 + \delta^{(n-1)}$ into the context construction and expanding yields the paper's Equation 7:

$c^{(t_c)}_{n-1} = \underbrace{(1 - \sigma_{t_c}) x^{(n-1)}_0}_{\text{true signal}} + \underbrace{(1 - \sigma_{t_c}) \delta^{(n-1)}}_{\text{propagated bias}} + \underbrace{\sigma_{t_c} \eta}_{\text{stochastic perturbation}}$

where $x^{(n-1)}_0$ is the unknown ground-truth clean latent, $\delta^{(n-1)}$ is the accumulated prediction error, $\eta$ is independent Gaussian noise, and $\sigma_{t_c}$ is the noise level at context time $t_c$.

What it computes: this decomposition separates the context into three additive components. The first term is the useful signal—the ground-truth content of the previous frame, scaled by $(1 - \sigma_{t_c})$. The second term is the harmful bias—the model's prediction error, scaled by exactly the same coefficient $(1 - \sigma_{t_c})$. The third term is independent stochastic noise, scaled by $\sigma_{t_c}$. The critical observation is that the true signal and the propagated bias share the same coefficient $(1 - \sigma_{t_c})$: any reduction in bias comes with an equal reduction in signal.

Why this form: the decomposition makes explicit the fundamental trade-off that the context noise level $t_c$ controls. Raising $t_c$ (adding more noise to the context) reduces both terms equally—the useful signal is attenuated, but so is the harmful error. Lowering $t_c$ (providing cleaner context) preserves the signal but also preserves the error. There is no free parameter that can reduce bias without reducing information. Standard AR methods implicitly choose $t_c = 0$, which maximizes signal but also maximizes bias propagation: $c^{(0)}_{n-1} = x^{(n-1)}_0 + \delta^{(n-1)}$, with no attenuation of either term.

The temporal causality constraint. The paper argues that not all values of $t_c$ are equally valid—there is a minimum information requirement for temporal coherence. Let $\text{SNR}(t) = (1 - \sigma_t)^2 / \sigma_t^2$ denote the signal-to-noise ratio at time $t$. After denoising step $j$, block $n$ is at noise level $t_{j+1}$ and therefore has SNR $\text{SNR}(t_{j+1})$. For the context to provide sufficient information to guide this block's generation, it must have at least as much signal relative to noise as the current block already possesses:

$\text{SNR}(t_c) \geq \text{SNR}(t_{j+1})$

where $\text{SNR}(t) = (1 - \sigma_t)^2 / \sigma_t^2$ is the signal-to-noise ratio of the forward process at time $t$, $t_c$ is the context noise level, and $t_{j+1}$ is the output noise level of the current denoising step.

What it computes: this inequality defines the set of valid context noise levels for step $j$. Any $t_c$ satisfying this constraint provides at least as much information about the true previous frame as the current block contains about its own true content after the denoising step. In other words, the context is not the information bottleneck—the current block's own noise level is.

Why this form: SNR is a natural measure of information content in the diffusion framework. Because $\text{SNR}(t)$ is strictly decreasing in $t$ (higher $t$ means more noise, lower SNR), the constraint $\text{SNR}(t_c) \geq \text{SNR}(t_{j+1})$ is equivalent to $t_c \leq t_{j+1}$. The constraint has a clear operational meaning: the context must be at least as denoised as the current block's state after step $j$, otherwise the model would be conditioning on noisier information than what it already has.

Deriving the optimal context noise level. Since the bias coefficient $(1 - \sigma_{t_c})$ decreases monotonically as $t_c$ increases, choosing any $t_c < t_{j+1}$ would transmit more prediction error than necessary without providing additional useful information—the SNR of the context would exceed the requirement, but the extra signal would come with proportionally extra bias. The optimal choice is therefore the boundary of the constraint—the noisiest level that still satisfies temporal causality:

$t_c^* = t_{j+1}$

where $t_{j+1}$ is the output noise level after completing denoising step $j$, i.e., the noise level that both the context block and the current block share after the step finishes.

What it computes: the optimal context noise level is simply the output noise level of the current denoising step—not the input level $t_j$ (which would be too noisy and violate causality), not the clean level $t = 0$ (which would maximize bias propagation), but exactly the level that the context block reaches after it undergoes the same denoising step. This is why the paper calls it "matched-noise context."

Why this form: this result is elegant because it removes a degree of freedom—there is no hyperparameter to tune for the context noise level. The optimal choice falls out directly from the structure of the problem: the bias–information trade-off is resolved by the temporal causality constraint, which picks out a unique noise level at each step. This contrasts with prior methods that implicitly or explicitly set $t_c = 0$, which can now be seen as a suboptimal corner of the feasible region.

What this means operationally. For the very first denoising step ($j = 1$), all blocks start at $t_1$ and are denoised to $t_2$. Block $B_1$ has no predecessor, so it is denoised unconditionally. Block $B_2$ conditions on $B_1$ at noise level $t_2$—the same level $B_2$ is being denoised to. Block $B_3$ conditions on $B_2$ at $t_2$, and so on. For the second step ($j = 2$), all blocks are denoised from $t_2$ to $t_3$, and each block $B_n$ conditions on $B_{n-1}$ at $t_3$. This pattern continues: at every step, the context is presented at exactly the noise level that both blocks will share after the step completes.

---

Hierarchical Denoising Inference Procedure

The generation grid. The inference procedure can be visualized as an $N \times S$ grid, where rows correspond to blocks $B_1, \ldots, B_N$ and columns correspond to denoising steps $j = 1, \ldots, S$. Each cell $(n, j)$ represents the state of block $n$ after denoising step $j$, at noise level $t_{j+1}$. The conventional block-first order fills this grid row by row (complete all $S$ steps for block 1, then all $S$ steps for block 2, etc.). The hierarchical denoising order fills the grid column by column (complete step 1 for all $N$ blocks, then step 2 for all $N$ blocks, etc.).

Algorithm 1 formal specification. The paper provides pseudocode as Algorithm 1. The procedure takes as input a noise schedule $t_1 > t_2 > \dots > t_S \approx 0$ (with $t_1 = 1$ as the initial pure-noise state) and initial noise tensors $\{x^{(n)}_{t_1}\}_{n=1}^N$ for all $N$ blocks, each drawn independently from $\mathcal{N}(0, I)$. It produces generated clean blocks $\{\hat{x}^{(n)}_0\}_{n=1}^N$.

The outer loop iterates over denoising steps $j = 1, \ldots, S$. Within each step, an inner loop iterates over blocks $n = 1, \ldots, N$ in causal order, maintaining a KV cache of previously processed blocks. For each block $n$ at step $j$, the update is:

$x^{(n)}_{t_{j+1}} = x^{(n)}_{t_j} + v_\theta\left(x^{(n)}_{t_j}, t_j \mid x^{(< n)}_{t_{j+1}}\right) \cdot (\sigma_{t_{j+1}} - \sigma_{t_j})$

where $x^{(n)}_{t_j}$ is the noisy latent of block $n$ at the start of step $j$, $v_\theta$ is the velocity-predicting neural network (the DiT backbone), $t_j$ is the current timestep embedding, $x^{(< n)}_{t_{j+1}}$ denotes the context from all previous blocks $1, \ldots, n-1$ at noise level $t_{j+1}$ (the output level of the current step), and $\sigma_{t_{j+1}} - \sigma_{t_j}$ is the step size in noise-space from the flow matching Euler update (Equation 3).

What it computes: this is a standard Euler integration step of the probability-flow ODE (Equation 3), but with a crucial modification to the conditioning: the velocity network $v_\theta$ receives context $x^{(< n)}_{t_{j+1}}$ at the output noise level $t_{j+1}$, not at the input noise level $t_j$ and not at the clean level $t = 0$. The velocity prediction moves block $n$ from noise level $t_j$ to $t_{j+1}$, and this prediction is informed by the previous blocks that have already been moved to $t_{j+1}$ within the same step. After the inner loop completes, the KV cache is updated with all blocks at their new noise level $t_{j+1}$, making them available as context for the next denoising step.

Why this form: the causal order within each step ($n = 1, 2, \ldots, N$) ensures that when block $B_n$ is being denoised, blocks $B_1, \ldots, B_{n-1}$ have already been processed at the current step and are available at noise level $t_{j+1}$. This satisfies the optimality condition $t_c = t_{j+1}$ derived above. The KV cache mechanism avoids recomputing attention keys and values for previously processed blocks, which would be computationally wasteful since their latents at $t_{j+1}$ are now fixed for the remainder of the step.

The KV cache update detail. After the inner loop over blocks completes for step $j$, the KV cache stores the key-value pairs computed from $\{x^{(n)}_{t_{j+1}}\}_{n=1}^N$. At the next step $j+1$, when denoising from $t_{j+1}$ to $t_{j+2}$, block $n$ will again condition on blocks $< n$, but now those blocks are at noise level $t_{j+2}$ after their own denoising. The keys and values from step $j$ at level $t_{j+1}$ are no longer needed—they are overwritten. This means the model conditions on context at a progressively cleaner noise level as denoising proceeds: at step 1, context is at $t_2$ (very noisy); at step $S-1$, context is at $t_S \approx 0$ (nearly clean). This progressive refinement of context mirrors the progressive refinement of the blocks themselves.

Initialization. Before the first step, all $N$ blocks are initialized with independent Gaussian noise: $x^{(n)}_{t_1} \sim \mathcal{N}(0, I)$ for $n = 1, \ldots, N$. There is no context for $B_1$ at any step (it is the first block in the sequence), so its denoising is unconditional with respect to previous blocks—it only conditions on any global conditioning signals (e.g., text prompts). The paper uses a sliding-window KV cache with a constant attention window of 5 seconds, meaning that blocks beyond a 5-second lookback are dropped from the cache, keeping memory usage bounded regardless of total video length.

Pipelined parallelism. The paper identifies a structural property of the $N \times S$ grid that enables significant wall-clock speedup. In Algorithm 1, cell $(n, j)$ depends on cells $(1, j), \ldots, (n-1, j)$ (the previous blocks at the same step, via causal attention) and cell $(n, j-1)$ (the same block at the previous step, the starting point for the Euler update). Cells on the same anti-diagonal of the grid—positions where $n + j$ is constant—are mutually independent because no cell on an anti-diagonal depends on any other cell on the same anti-diagonal. For example, cell $(1, 2)$ and cell $(2, 1)$ are independent: $(1, 2)$ depends on $(1, 1)$, while $(2, 1)$ depends on $(1, 1)$ and $(2, 0)$ (which does not exist), but they do not depend on each other.

This independence allows the $N + S - 1$ anti-diagonals to be processed in parallel with pipelining. The paper assigns each denoising step to a dedicated process. Blocks flow through the pipeline: process 1 handles step 1 for all blocks, process 2 handles step 2 for all blocks, etc. Inter-stage latents are exchanged via asynchronous point-to-point communication. When process $j$ finishes block $n$, it sends $x^{(n)}_{t_{j+1}}$ to process $j+1$ (which needs it as input for the next step) and to any other processes that need it as context.

KV cache fusion optimization. A naive implementation of the inner loop would require two forward passes per block: one to compute and cache the keys/values for block $B_n$ at $t_{j+1}$ (updating the KV cache), and another to denoise block $B_{n+1}$ attending to that cache. This would cost $2N$ forward passes per step.

The paper observes that under causal attention, these two operations can be fused into a single forward pass by concatenating the context block and the target block along the frame dimension, with different per-frame timesteps. Specifically, to process block $B_{n+1}$ at step $j$, the input is formed by concatenating $[x^{(n)}_{t_{j+1}}, x^{(n+1)}_{t_j}]$ with per-frame timesteps $[t_{j+1}, \ldots, t_{j+1}, t_j, \ldots, t_j]$. Under causal attention, the first segment (context block $n$ at $t_{j+1}$) writes its keys and values into the cache while attending to previous blocks; the second segment (block $n+1$ at $t_j$) attends to all previous blocks including the freshly written block $n$, and produces velocity predictions that advance it to $t_{j+1}$. This fused operation costs one forward pass instead of two.

With this fusion, the per-step cost reduces to $N + 2$ forward passes: one standalone denoising pass for block $B_1$ (which has no predecessor context to fuse with), $N - 1$ fused passes for blocks $B_2$ through $B_N$, and one trailing cache-write pass to update the KV cache with the final block's states for use in the next denoising step. The paper reports this yields a $\sim 1.8\times$ wall-clock speedup in their 4-step setting ($S = 4$), translating to 30 fps throughput and 0.30 s per-chunk latency compared to 17 fps and 0.69 s for other distilled AR models with the same backbone (Table 1).

---

Training with Self-Rollout Under Hierarchical Denoising

Why retraining is necessary. The paper demonstrates empirically (Table 3, "w/o re-training" row) that applying the hierarchical denoising procedure at inference time to a model trained under the conventional block-first paradigm produces poor results. Specifically, Quality drops from 0.846 to 0.767 and Semantic drops from 0.723 to 0.559. While drift does improve (0.309 vs. 0.355 for Self-Forcing), the severe degradation in visual quality makes this approach impractical. The problem is a train–test mismatch: the model was trained with clean or nearly-clean context and has never encountered the matched-noise context that hierarchical denoising provides at inference time. The model must be retrained from scratch under the hierarchical schedule so that the training-time context noise distribution matches the inference-time distribution.

Training data preparation. The paper follows the Self-Forcing (Anonymous, 2025) recipe for data generation. The base model is Wan2.1-1.3B (Team, 2025), a bidirectional-attention DiT pretrained on large-scale video data. The authors sample 16,000 ODE solution pairs from this base model—each pair consists of a noisy latent at some timestep and the corresponding clean latent obtained by running the full ODE integration. These pairs serve as the distillation targets. All training is done on 5-second video clips, which are partitioned into blocks (chunks) of 3 latent frames each.

Self-rollout procedure. During each training iteration, the model (the "student" $v_\theta$) performs a full hierarchical rollout to generate context for the next block. For a training video consisting of multiple blocks, the procedure mirrors inference:

1. Initialize all blocks with independent noise at $t_1$.
2. For step $j = 1$, denoise $B_1$ unconditionally to $t_2$, then denoise $B_2$ conditioned on the generated $B_1$ at $t_2$, then $B_3$ conditioned on $B_2$ at $t_2$, etc.
3. Repeat for steps $j = 2, \ldots, S$.

The generated block $B_n$ is then used as context for training the denoising of block $B_{n+1}$. This is the key difference from teacher forcing: the context comes from the model's own (current) parameters, not from ground-truth data.

The DMD reverse-KL objective. The primary training loss is the Distribution Matching Distillation objective from Self-Forcing, formulated as a reverse KL divergence:

$\mathcal{L}_{\text{DMD}} = \mathbb{E}_{t, x_t}\left[D_{\text{KL}}\left(p_\theta(x_0 \mid x_t) \,\|\, p_{\text{teacher}}(x_0 \mid x_t)\right)\right]$

where $p_\theta(x_0 \mid x_t)$ is the distribution over clean samples induced by the student's single Euler step from noisy latent $x_t$, $p_{\text{teacher}}(x_0 \mid x_t)$ is the distribution obtained by running the teacher model for many ODE steps from the same $x_t$, $D_{\text{KL}}$ is the Kullback-Leibler divergence, and the expectation is taken over diffusion timesteps $t$ and noisy latents $x_t$.

What it computes: this loss measures how different the student's one-step denoising distribution is from the teacher's full multi-step denoising distribution, averaged over all noise levels. For each training sample, a noisy latent $x_t$ is produced (by adding noise to a clean latent according to the schedule), the student takes one Euler step to predict $\hat{x}_0$, the teacher takes many steps to produce a reference $x_0^{\text{ref}}$, and the KL divergence between the resulting distributions is estimated using a learned critic network that approximates the score difference. The student's parameters are updated to minimize this divergence.

Why this form: reverse KL, $D_{\text{KL}}(p_\theta \| p_{\text{teacher}})$, is mode-seeking—it penalizes the student heavily for generating samples that are unlikely under the teacher (the teacher's density is in the denominator inside the log), but penalizes only weakly for failing to cover all of the teacher's modes (if the teacher has mass in a region the student ignores, the contribution to $D_{\text{KL}}$ is small because $p_\theta$ multiplies the log-ratio and $p_\theta \approx 0$ in that region). This property makes reverse KL good for producing high-quality, sharp samples (the student is forced to stay in high-probability regions) but creates a risk of mode collapse: the student can achieve low loss by concentrating all its probability mass on a single high-density mode of the teacher, ignoring other modes.

Implementation details. The paper uses Wan2.1-14B as the teacher model for the DMD critic—a substantially larger model that produces higher-quality multi-step denoising trajectories. The critic model and generator (student) are updated at a 5:1 ratio, meaning the critic receives five updates for every one generator update, keeping the critic's score estimates accurate as the generator distribution shifts. The learning rate is $2 \times 10^{-6}$ with a total batch size of 64, trained for 20,000 steps on 5-second clips.

The low-motion shortcut problem. As training progresses, the paper observes a specific failure mode of the reverse-KL objective in the hierarchical denoising setting: motion diversity progressively collapses, with the Dynamic score (a VBench sub-metric measuring motion magnitude and variety) dropping drastically. Table 3 shows that without the forward-KL regularizer ("w/o LFKL" row), dynamics collapse to 0.445, compared to 0.686 for the full HiAR model and 0.690 for the bidirectional Wan2.1 teacher.

The root cause, as the paper explains, is the interplay between two factors:

1. Mode-seeking reverse KL. The student can reduce $\mathcal{L}_{\text{DMD}}$ by generating outputs that are inherently easier to denoise consistently—low-motion, near-static videos where consecutive frames are almost identical. Such videos produce smaller rollout errors because there is little change to predict, and the DMD loss decreases because the student's single-step predictions are more reliably close to the teacher's multi-step predictions when the target is static.

2. Hierarchical denoising amplifies the effect. Conditioning on contexts at varying noise levels (rather than always-clean context) makes the denoising task more difficult for the student—it must learn to extract temporal information from partially noisy previous frames. This increased difficulty requires more training steps to achieve good visual quality. More training steps give the mode-seeking objective more iterations to collapse onto the low-motion attractor.

The result is a "low-motion shortcut": the model learns that the easiest way to minimize the DMD loss is to produce videos where nothing moves, which is a degenerate solution for a video generation system.

---

Forward-KL Regularization via Trajectory Distillation

Motivation and intuition. To counteract the mode-seeking tendency of reverse KL, the paper introduces an auxiliary loss that encourages the opposite behavior: mode-covering, where the student is penalized for ignoring regions of the teacher's output distribution. The natural choice is forward KL, $D_{\text{KL}}(p_{\text{teacher}} \| p_\theta)$, which penalizes the student heavily when the teacher has high probability in a region where the student has low probability (the teacher's density multiplies the log-ratio). Forward KL encourages the student to spread its mass across all of the teacher's modes, preserving diversity at the potential cost of occasionally generating lower-quality samples.

Directly estimating forward KL for high-dimensional video distributions is intractable. The paper instead uses a trajectory matching approach: run the teacher through a long, high-quality denoising trajectory, extract checkpoints at the student's coarse schedule, and supervise the student to match each teacher step with a single Euler step.

Teacher trajectory generation. The paper samples 20,000 denoising trajectories from the Wan2.1-1.3B base model (the same architecture as the student, but run with many steps). Each trajectory uses 50 ODE steps to go from pure noise to a clean video. From each trajectory, the authors extract $S = 4$ checkpoints aligned with the student's 4-step schedule at noise levels $t_1 > t_2 > t_3 > t_4 \approx 0$, yielding reference latents $\{x^{\text{ref}}_{t_1}, x^{\text{ref}}_{t_2}, x^{\text{ref}}_{t_3}, x^{\text{ref}}_{t_4}\}$. The first checkpoint $x^{\text{ref}}_{t_1}$ is pure noise (the starting point), and the last checkpoint $x^{\text{ref}}_{t_4}$ is the clean generated video.

Forward-KL loss formulation. For each consecutive pair of reference checkpoints $(x^{\text{ref}}_{t_i}, x^{\text{ref}}_{t_{i+1}})$, the student is trained to predict the velocity that would move from the first to the second in a single Euler step:

$\mathcal{L}_{\text{FKL}} = \mathbb{E}_i\left[\left\| v_\theta(x^{\text{ref}}_{t_i}, t_i) - \frac{x^{\text{ref}}_{t_{i+1}} - x^{\text{ref}}_{t_i}}{\sigma_{t_{i+1}} - \sigma_{t_i}} \right\|^2\right]$

where $i$ indexes the denoising steps, $v_\theta(x^{\text{ref}}_{t_i}, t_i)$ is the student's predicted velocity at the reference latent and timestep, $\frac{x^{\text{ref}}_{t_{i+1}} - x^{\text{ref}}_{t_i}}{\sigma_{t_{i+1}} - \sigma_{t_i}}$ is the ground-truth velocity implied by the teacher's trajectory (the finite-difference approximation of the true velocity field between the two checkpoints), and $\|\cdot\|^2$ is the squared Euclidean norm.

What it computes: this is a simple mean squared error (MSE) between the student's velocity prediction and the teacher's implied velocity, averaged over denoising steps and trajectory samples. For each step $i$ in the student's schedule, the target velocity is computed as the normalized displacement between consecutive teacher checkpoints. The student is trained to reproduce this displacement in a single step, which amounts to learning the teacher's probability flow.

Why this form: minimizing MSE between the student's velocity and the teacher's implied velocity corresponds to maximizing the log-likelihood of the teacher's trajectory under a Gaussian model—this is equivalent to minimizing a forward KL divergence between the student's transition distribution and the teacher's (approximately) delta distribution at the trajectory point. Because the targets $x^{\text{ref}}_t$ are drawn from the teacher's marginal distribution at each timestep, the expectation over trajectories covers the teacher's output modes. The squared-error loss penalizes the student for large deviations from any teacher trajectory point, encouraging it to cover all modes present in the teacher samples, not just the highest-density ones.

This is in contrast to the DMD loss, which uses a learned critic to estimate the score difference and does not directly supervise the student's output to match specific teacher samples. The DMD loss only cares about the distribution matching, allowing the student to ignore modes as long as the ignored regions have low $p_\theta$ density. The forward-KL loss cares about pointwise matches to teacher trajectories, penalizing the student for failing to reproduce any teacher sample.

Design choice 1: bidirectional-attention mode only. A critical design decision is that $\mathcal{L}_{\text{FKL}}$ is computed exclusively in bidirectional-attention mode, not in causal mode. This means the student processes the full video with bidirectional (non-causal) attention when computing the forward-KL loss, seeing all frames simultaneously just as the teacher does. The DMD loss ($\mathcal{L}_{\text{DMD}}$) continues to be computed in causal mode during self-rollout training, maintaining train–test consistency for the autoregressive inference path.

The justification for this decoupling comes from an empirical observation shown in Figure 4: the Dynamic scores under bidirectional and causal attention during training are strongly positively correlated (Pearson $r = 0.968$, $p < 10^{-6}$). Both attention modes exhibit a consistent decline in dynamics over training when $\mathcal{L}_{\text{FKL}}$ is absent, and the correlation means that intervening to preserve bidirectional dynamics effectively preserves causal dynamics as well.

Why not compute $\mathcal{L}_{\text{FKL}}$ directly in causal mode? The paper's ablation (Table 3, "causal + 1 step") shows this performs worse: Dynamic drops to 0.625 (vs. 0.686 for bidirectional mode) and Quality drops to 0.828 (vs. 0.846). The explanation, illustrated in Figure 5, is that causal and bidirectional denoising produce qualitatively different per-step outputs. Under bidirectional attention, all frames receive symmetric treatment and exhibit uniform quality and blur at each denoising step. Under causal attention, frames become progressively sharper along the temporal axis: earlier frames fix low-frequency structure first, reducing uncertainty for later frames, which can then concentrate on higher-frequency details. This means the teacher's bidirectional trajectory targets are mismatched with the student's causal-mode outputs—the teacher's uniform refinement pattern does not align with the student's progressive sharpening pattern. By restricting $\mathcal{L}_{\text{FKL}}$ to bidirectional mode, the paper regularizes global motion dynamics using well-matched targets without interfering with the causal DMD loss that handles the autoregressive generation pathway.

Design choice 2: early-step restriction. The paper restricts $\mathcal{L}_{\text{FKL}}$ to only the first $K$ of $S$ denoising steps, with the default being $K = 1$ for their 4-step schedule. The rationale is that motion dynamics—the large-scale movements of objects, camera motion, and scene changes—are governed by low-frequency structures established during the earliest denoising steps, when the video transitions from pure noise to coarse structure. Later denoising steps refine high-frequency details (textures, edges, fine motion) but do not substantially alter the overall motion pattern.

The ablation in Table 3 confirms this: increasing $K$ from 1 to 2 or 4 yields marginal gains in dynamics (0.693 and 0.691 vs. 0.686 for $K = 1$) but monotonically degrades quality and increases drift. At $K = 2$, Quality drops to 0.835 and drift rises to 0.296; at $K = 4$, Quality falls further to 0.813 and drift to 0.306. Constraining later steps interferes with the model's denoising capacity—high-frequency refinement benefits from the student learning its own distribution rather than being forced to match the teacher's specific trajectory—while providing diminishing returns for motion diversity.

Overall training objective. The final loss is a weighted combination:

$\mathcal{L} = \mathcal{L}_{\text{DMD}} + \lambda \mathcal{L}_{\text{FKL}}$

where $\mathcal{L}_{\text{DMD}}$ is the reverse-KL distribution matching loss computed via the learned critic in causal self-rollout mode, $\mathcal{L}_{\text{FKL}}$ is the forward-KL trajectory matching loss computed in bidirectional-attention mode on the first $K = 1$ denoising step only, and $\lambda = 0.1$ is a balancing weight that determines the relative strength of the motion-diversity regularizer.

What it computes: at each training step, the model receives two gradient signals. The first, from $\mathcal{L}_{\text{DMD}}$, pushes the student to produce outputs that the critic (trained adversarially against the teacher) judges as high-quality—this is the primary quality driver. The second, from $\lambda \mathcal{L}_{\text{FKL}}$, pushes the student to reproduce the teacher's velocity field on the first denoising step in bidirectional mode—this is the motion-diversity preservative, preventing collapse onto static outputs.

Why this form: the two losses serve complementary purposes that a single loss cannot achieve. Reverse KL alone produces high-quality but low-diversity outputs (mode collapse). Forward KL alone would produce diverse but potentially lower-quality outputs (mode covering can admit samples from low-density regions). The combination, with $\lambda$ controlling the trade-off, allows the model to maintain both high quality (from DMD) and high motion diversity (from FKL). The bidirectional/causal decoupling prevents gradient interference: the two losses operate in different attention modes and therefore affect the model's parameters through partially distinct computational pathways, reducing the risk that the forward-KL regularization degrades the causal generation quality.

Training hyperparameters summary. The paper provides these specific values: base model Wan2.1-1.3B, teacher for critic Wan2.1-14B, training data 16,000 ODE solution pairs from the base model plus 20,000 trajectory samples for $\mathcal{L}_{\text{FKL}}$, denoising steps $S = 4$, forward-KL steps $K = 1$, forward-KL weight $\lambda = 0.1$, chunk size 3 latent frames per block, learning rate $2 \times 10^{-6}$, batch size 64, critic-to-generator update ratio 5:1, total training steps 20,000, training clip duration 5 seconds, sliding-window KV cache with 5-second attention window at inference.

---

Architecture and Attention Masking

Causal attention masking. The autoregressive property is enforced through attention masking in the DiT backbone. During the self-rollout and inference procedures, each block $B_n$ attends only to blocks $B_1, \ldots, B_n$ (itself and previous blocks), not to future blocks $B_{n+1}, \ldots, B_N$. This is implemented via a causal mask in the self-attention layers: queries from block $n$ can attend to keys and values from blocks $\leq n$, and the attention weights for blocks $> n$ are set to $-\infty$ before the softmax. This ensures the model cannot "cheat" by looking ahead at future frames, which would make the training trivial but the inference impossible (since future frames are not yet generated at inference time).

Bidirectional attention (for forward-KL only). When computing $\mathcal{L}_{\text{FKL}}$, the causal mask is removed and replaced with full bidirectional attention—every frame attends to every other frame. This is the standard attention mode of the pretrained Wan2.1 base model, and it allows the teacher trajectory matching to operate without the sequential constraints of autoregressive generation. The model's weights are shared between causal and bidirectional modes; the only difference is the attention mask.

Sliding-window KV cache. For very long videos (well beyond the 20-second evaluation in the paper), maintaining a KV cache of all previously generated blocks would cause memory usage to grow linearly with video duration. The paper uses a sliding window of 5 seconds: only blocks within the most recent 5 seconds are retained in the KV cache; older blocks are evicted. This means the model has a finite temporal context window, analogous to the context length in language models, and must generate coherent continuations from this limited history. The 5-second window is chosen to match the training clip duration, so the model never encounters a longer context at inference than it saw during training.

Chunk structure. All methods operate in a "chunk-wise" manner where each block (chunk) contains 3 latent frames. The latent frames are produced by the VAE encoder of the Wan2.1 model, which compresses the raw video frames into a lower-dimensional latent space. The denoising operates entirely in this latent space, and the final generated latents are decoded back to pixel space by the VAE decoder. The 3-frame chunk size is a design choice that balances context granularity (finer chunks allow more frequent conditioning updates) against computational efficiency (coarser chunks reduce the number of autoregressive steps).

---

Summary of Design Choices and Their Justifications

• Matched-noise context ($t_c = t_{j+1}$) over clean context ($t_c = 0$): derived analytically from the bias–information trade-off. Clean context maximizes signal but also maximizes error propagation; matched-noise context attenuates both equally, preserving temporal causality while reducing inter-block error transmission.

• Step-first generation order over block-first order: enables the matched-noise conditioning—you cannot provide context at $t_{j+1}$ if previous blocks haven't been denoised to $t_{j+1}$ yet. The step-first order ensures that when block $n$ is processed at step $j$, block $n-1$ has already reached $t_{j+1}$.

• Retraining under hierarchical schedule over inference-only application: empirical necessity—the train–test mismatch from applying hierarchical denoising to a block-first-trained model causes severe quality degradation (Quality drops from 0.846 to 0.767).

• Self-rollout training over teacher forcing: closes the train–test gap by using the model's own generated outputs as context during training, matching the inference-time context distribution.

• Forward-KL regularizer over reverse-KL-only training: prevents motion collapse (Dynamic drops from 0.686 to 0.445 without FKL). Forward KL encourages mode-covering, counteracting the mode-seeking tendency of reverse KL.

• Bidirectional-attention mode for FKL over causal mode: strong correlation between bidirectional and causal dynamics (Pearson $r = 0.968$) means bidirectional regularization effectively constrains causal dynamics. Bidirectional mode also provides better-matched targets since the teacher's trajectory uses bidirectional attention and produces uniform frame quality.

• First-step-only FKL ($K = 1$) over multi-step: motion dynamics are governed by low-frequency structure established in the earliest denoising steps. Constraining later high-frequency refinement steps degrades quality without meaningfully improving motion diversity.

• Pipelined parallelism with KV cache fusion over sequential execution: exploits the independence of cells on the same anti-diagonal of the $N \times S$ grid, achieving $\sim 1.8\times$ speedup through parallel processing and reducing forward passes per step from $2N$ to $N + 2$ via fused context-and-denoise operations.

4. Key Insights and Innovations

Innovation 1: The Matched-Noise Context Principle as a Principled Resolution of the Bias–Information Trade-Off

The paper's most conceptually distinctive contribution is not the hierarchical denoising architecture itself, but the analytic principle that justifies it: the idea that the optimal noise level for context in autoregressive video diffusion is exactly the output noise level of the current denoising step ($t_c^* = t_{j+1}$), and that this optimum arises from a fundamental bias–information trade-off that prior methods failed to recognize, let alone resolve optimally.

What the field assumed before this work. The dominant assumption across all prior autoregressive video generation methods—from teacher forcing (Williams and Zipser, 1989; Gao et al., 2024) through Diffusion Forcing (Chen et al., 2024; Yin et al., 2025b) to Self-Forcing (Anonymous, 2025)—was that cleaner context is better context. The reasoning was intuitive and seemingly unassailable: the more precisely you show the model what came before, the better it can produce a coherent continuation. Self-Forcing, the immediate predecessor to this work, went to substantial lengths to close the train–test gap while preserving this assumption, generating clean-context predictions through self-rollout training but still feeding them to subsequent blocks at noise level $t_c = 0$.

What the paper reframes. HiAR identifies that this "cleaner-is-better" intuition contains an unexamined fallacy: in the autoregressive setting, context carries not only signal but also accumulated prediction error, and the clean-context choice amplifies both equally. The insight comes from recognizing that the context noise level $t_c$ controls a single shared coefficient $(1 - \sigma_{t_c})$ that multiplies both the true signal $x^{(n-1)}_0$ and the prediction error $\delta^{(n-1)}$ in the context construction (Equation 7). There is no way to attenuate the error without attenuating the signal by exactly the same factor. The field had implicitly optimized for maximum signal (set $t_c = 0$) without accounting for the cost in error propagation.

The paper then introduces a critical second piece: a temporal causality constraint (Equation 8) that establishes a lower bound on how noisy the context can be while still providing sufficient information. The constraint $\text{SNR}(t_c) \geq \text{SNR}(t_{j+1})$ formalizes the idea that context must be at least as informative as the current block's own state after the denoising step. This constraint carves out the feasible region, and the optimal point is at its boundary—the noisiest context that still satisfies temporal causality, which is exactly $t_c = t_{j+1}$.

Why this is a conceptual advance, not just a design tweak. This is not an incremental improvement on Self-Forcing; it is a fundamental reframing of what context should be in autoregressive diffusion. Before this work, the context noise level was effectively a hidden hyperparameter that everyone set to 0 without justification. After this work, it is a principled quantity derived from the structure of the problem, with a closed-form optimum that varies per denoising step. The fact that the optimum is dynamic—$t_c$ decreases as denoising progresses, since $t_{j+1}$ decreases with $j$—means that the "right" context is not just "noisier than clean" but "matched to the current refinement stage," a nuance that no prior method captured.

This principle has independent significance beyond video generation: any autoregressive diffusion process where earlier outputs serve as context for later outputs potentially faces the same bias–information trade-off, and the matched-noise derivation provides a general template for resolving it. The paper does not explore this generalization, but the framing is portable.

Evidence anchor. The ablation in Table 2 directly tests the three points on the feasible spectrum: $t_c = t_j$ (input noise level, below the causality constraint), $t_c = t_{j+1}$ (the derived optimum), and $t_c = 0$ (the prior convention). The middle setting achieves the best Quality (0.846) and Semantic (0.723) while reducing Drift by 27.6% relative to $t_c = 0$ (0.257 vs. 0.355), confirming the theoretical prediction.

Innovation 2: The Generation Order as a First-Class Architectural Degree of Freedom

The paper identifies that the conventional "block-first" generation order in autoregressive diffusion—fully denoise block 1, then fully denoise block 2, etc.—is not a logical necessity but a design choice with profound consequences for error propagation. By inverting this to a "step-first" order (one denoising step for all blocks, then the next step for all blocks), HiAR makes the matched-noise conditioning principle implementable and exposes a new axis of architectural design that the field had not previously considered as tunable.

What the field assumed before this work. The block-first order was so deeply ingrained in autoregressive video generation that it was essentially invisible as a design choice. It followed naturally from the autoregressive paradigm borrowed from language modeling: generate token 1, then token 2 conditioned on token 1, then token 3 conditioned on tokens 1–2, etc. Translating this to video diffusion meant: generate block 1 completely (through all denoising steps), then generate block 2 completely conditioned on the fully generated block 1, and so on. This order seemed intrinsic to the very definition of "autoregressive"—first things first.

What HiAR reveals. The paper demonstrates that "autoregressive" specifies a causal dependency structure (block $n$ depends on blocks $< n$), not a specific execution order. As long as the dependency graph is respected—block $n$ at step $j$ depends on blocks $< n$ at step $j$ and block $n$ at step $j-1$—the actual traversal order through the $N \times S$ computation grid is a degree of freedom. HiAR exploits this freedom by traversing column-by-column (step-first) rather than row-by-row (block-first).

This is more than a reordering for its own sake. The step-first traversal is what makes matched-noise conditioning possible: you cannot provide context at noise level $t_{j+1}$ for block $B_n$ unless block $B_{n-1}$ has already been denoised to $t_{j+1}$, which requires processing block $B_{n-1}$ at step $j$ before processing block $B_n$ at step $j$. In a block-first order, by the time you start block $B_n$, block $B_{n-1}$ is already at $t_S \approx 0$ (fully denoised), and the intermediate noise levels are gone—there is no $t_{j+1}$ context available. The conventional order forces the clean-context convention; the step-first order liberates the context noise level as a tunable quantity.

Parallelism as a derived property, not a separate contribution. The paper presents the $\sim 1.8\times$ speedup from pipelined parallelism as a benefit of hierarchical denoising, but the deeper insight is that the step-first traversal exposes parallelism that the block-first order structurally conceals. In the block-first order, there is no parallelism across blocks because block $n$ cannot begin until block $n-1$ finishes all $S$ steps. In the step-first order, cells on anti-diagonals of the $N \times S$ grid are independent, enabling pipelined execution. The speedup is not an optimization trick applied on top of the architecture; it is a property that emerges from the reordered computation graph. This suggests a broader principle: generation order choices in autoregressive models carry implicit parallelism consequences that deserve explicit analysis.

Why this is fundamental, not incremental. The generation order is not a hyperparameter—it changes what information is available at each computation step and therefore what the model can learn. The paper shows that applying step-first generation at inference time to a block-first-trained model causes severe quality degradation (Table 3, "w/o re-training": Quality drops to 0.767 from 0.846). This means the generation order is baked into the model's learned representations during training; you cannot simply swap orders post-hoc. By surfacing this as a first-class design dimension, the paper opens a line of inquiry that extends beyond the specific matched-noise solution: what other generation orders might yield different information-flow properties, and what training procedures would they require?

Evidence anchor. The comparison between HiAR (trained with step-first order) and "w/o re-training" in Table 3 demonstrates that the generation order is a training-time commitment, not an inference-time switch. The full HiAR with retraining achieves 0.846 Quality; the same architecture with hierarchical denoising applied only at inference achieves 0.767, confirming that the model's internal representations are shaped by the generation order it was trained under.

Innovation 3: Forward-KL Regularization in a Decoupled Attention Mode as a Solution to Reverse-KL Mode Collapse in Video Distillation

The paper's third distinctive contribution is a diagnostic insight into why distillation-based autoregressive video generation suffers motion collapse, coupled with a regularization strategy—forward-KL trajectory matching in bidirectional-attention mode only—that addresses the root cause without interfering with the primary distillation objective.

The diagnostic insight. Prior work using reverse-KL objectives for diffusion distillation (Self-Forcing, CausVid) observed that motion diversity degrades with training, but the mechanism was not clearly articulated. HiAR provides a precise diagnosis: the reverse-KL objective $D_{\text{KL}}(p_\theta \| p_{\text{teacher}})$ is mode-seeking, and in the video domain, the "easiest" high-density mode is low-motion or near-static output. This is not a generic mode-collapse problem—it is specific to the video setting because motion introduces complexity that makes consistent autoregressive denoising harder. A static video where frame $t$ is almost identical to frame $t+1$ is trivially easy to denoise: the velocity field is nearly zero everywhere, the rollout errors are minimal, and the DMD loss approaches its minimum. The model discovers this "shortcut" through normal gradient-based optimization, and hierarchical denoising amplifies the effect because conditioning on multi-level noisy contexts increases the difficulty of producing coherent motion, making the low-motion shortcut relatively more attractive.

This diagnostic is significant because it explains why prior methods plateau or degrade with extended training: they are not failing to optimize the loss; they are optimizing it too well, converging to a degenerate solution that the loss function cannot distinguish from genuine high-quality video. The Dynamic score of the full HiAR (0.686) nearly matches the teacher Wan2.1-1.3B (0.690), while Self-Forcing drops to 0.542, confirming that the collapse is a training artifact, not an inherent limitation of few-step AR generation.

The decoupled regularization strategy. The solution—adding a forward-KL loss in bidirectional-attention mode only—is novel not for the forward-KL concept itself (which is standard in generative modeling) but for two design choices that make it work in this specific context:

1. Attention-mode decoupling. The paper empirically discovers that motion dynamics under bidirectional and causal attention are strongly correlated (Pearson $r = 0.968$, Figure 4) during training without regularization. This correlation is not obvious a priori—one might expect that causal constraints would produce fundamentally different dynamics patterns than bidirectional attention. The strong correlation justifies a non-trivial design choice: compute the computationally and conceptually simpler forward-KL loss in bidirectional mode (where targets from the teacher's trajectory are well-matched, since the teacher also uses bidirectional attention) and trust that the regularization transfers to causal mode through the shared model parameters. This is a discovery about the relationship between attention modes, not merely a trick.

2. Early-step restriction. The finding that constraining only the first denoising step ($K = 1$) is sufficient—and that constraining more steps actually degrades performance—reveals that motion diversity is governed by low-frequency structure established in the earliest denoising stage. Later steps refine details but do not substantially alter the global motion pattern. This is an insight about the diffusion process itself: the denoising trajectory separates into a "layout" phase (early steps, where motion patterns are determined) and a "detailing" phase (later steps, where texture and sharpness are refined). Regularization is beneficial in the layout phase but harmful in the detailing phase, where precise matching to teacher trajectories constrains the model's ability to produce sharp, high-frequency content.

Why this is fundamental. This contribution addresses what would otherwise be a fatal limitation of the entire hierarchical denoising approach. Without the forward-KL regularizer, HiAR would produce stable, drift-free videos that hardly move (Dynamic 0.445 in Table 3)—hardly a compelling alternative to bidirectional models that produce rich motion within their fixed window. The regularizer is not an optional add-on; it is what makes the hierarchical denoising framework produce useful video rather than merely stable video. The fact that the solution requires understanding the correlation structure between attention modes, the phase separation in the denoising trajectory, and the interaction between regularization and the primary loss elevates this from an engineering fix to an insight about the system's dynamics.

Evidence anchor. Table 3 provides the key ablation: without LFKL, Dynamic collapses to 0.445 while Quality and Drift remain competitive (0.839 and 0.218 respectively), demonstrating that the mode collapse is a specific failure of motion diversity, not a general quality degradation. The bidirectional-mode correlation evidence in Figure 4 (Pearson $r = 0.968$) directly supports the attention-mode decoupling strategy.

Innovation 4: Drift as a Quantifiable, Perceptually-Grounded Metric for Long-Horizon Video Stability

The paper introduces a drift metric suite that measures temporal degradation not through aggregate quality scores (which can mask progressive decline by averaging over the full video) but through the rate of change of perceptual and statistical properties across temporal segments. This is a methodological contribution that addresses a gap in how the field evaluates long video generation.

What the field lacked before this work. Standard video generation benchmarks like VBench evaluate overall quality and semantic alignment but are not designed to capture temporal degradation profiles. A video that starts at high quality and slowly degrades over 20 seconds can achieve a similar aggregate VBench score to one that maintains moderate but consistent quality throughout—the average masks the trajectory. Methods that claim to solve the drift problem cannot be rigorously compared without a metric that directly quantifies drift. The field had qualitative observations of degradation (oversaturation, hue shifts, motion repetition) but no standardized quantitative measure.

What the drift metric captures. The metric divides each 20-second video into five equal temporal segments and computes per-segment statistics spanning multiple dimensions: perceptual quality (MUSIQ, CLIP-IQA), temporal coherence (DINOv2 consecutive-frame similarity, LPIPS consecutive-frame distance), and low-level statistics (HSV saturation mean, Laplacian variance as a sharpness proxy). For each of these per-segment statistic series, a linear fit is computed over the five segments, and the slope captures the rate of drift—how quickly the property degrades as the video progresses. The per-metric slopes are normalized and aggregated via weighted sum into a single Drift Score where lower values indicate better temporal stability.

This design has several thoughtful properties. Using slope rather than endpoint difference makes the metric robust to segment-level noise while capturing the overall trend. Using multiple complementary dimensions (perceptual quality, temporal coherence, low-level color and sharpness statistics) ensures that the metric captures different failure modes—a model might maintain sharpness but lose color fidelity, or maintain color but suffer motion collapse—and the weighted aggregation allows tuning sensitivity to different degradation types. The normalization across metrics with different natural scales makes the weighted sum meaningful.

Why this matters beyond this paper. The drift metric is not specific to HiAR; it is a general evaluation tool for any long-horizon video generation method. By providing a standardized, quantitative drift measure, it enables rigorous comparison between methods that make claims about temporal stability—a comparison that was previously limited to qualitative assessment or aggregate metrics that conflate average quality with temporal consistency. The paper's own use of the metric in Table 1 demonstrates its discriminatory power: CausVid shows the highest drift (0.842), consistent with its visible color degradation; Self-Forcing (0.355) and Causal Forcing (0.615) show intermediate drift; HiAR achieves the lowest (0.257). The 27.6% reduction in drift relative to Self-Forcing is a quantifiable claim that would be impossible to make rigorously with only VBench aggregate scores.

Evidence anchor. Table 1 reports the Drift metric for all compared methods, showing HiAR's 0.257 as the lowest among distilled AR models. The qualitative results in Figure 3 visually corroborate the metric: CausVid's high drift score (0.842) corresponds to severe green/yellow tint drift visible in the sampled frames, while HiAR's low drift score corresponds to stable color and detail across all segments.

5. Experimental Analysis

Evaluation Methodology

• Dataset. All experiments use the VBench benchmark (Huang et al., 2024), which measures 16 dimensions grouped into a Quality score and a Semantic score, providing comprehensive assessment of average generation quality. All models are sampled to generate 20-second videos to evaluate long-video capability. The paper also introduces a custom drift metric suite (described below in Metrics) specifically designed for long-horizon evaluation, computed by dividing each 20-second video into five temporal segments and measuring per-segment statistics.

• Base model(s). HiAR uses Wan2.1-1.3B (Team, 2025) as the backbone—a pretrained bidirectional-attention video diffusion transformer. The choice is deliberate: this model family serves as the foundation for all distilled autoregressive baselines (Self-Forcing, CausVid, Causal Forcing), enabling direct architecture-controlled comparison. For the DMD critic in distillation, the paper uses Wan2.1-14B as the teacher model, a substantially larger variant that produces higher-quality multi-step denoising trajectories. The forward-KL trajectory data is sampled from Wan2.1-1.3B itself (20,000 trajectories, 50 ODE steps each).

• Metrics. Three categories of metrics are reported:

  • VBench scores (0–1 scale): Total score aggregates all dimensions; Quality score covers visual fidelity dimensions; Semantic score covers text-alignment and content dimensions; Dynamic score specifically measures motion magnitude and variety. All models are evaluated at 20 seconds to assess long-video stability.
  • Drift Score (lower is better): A custom metric suite designed for long-horizon evaluation. Each 20-second video is divided into five equal temporal segments. Per-segment statistics include perceptual quality (MUSIQ, CLIP-IQA), temporal coherence (DINOv2 consecutive-frame cosine similarity, LPIPS consecutive-frame distance), and low-level statistics (HSV saturation mean, Laplacian variance as sharpness proxy). For each metric, a linear fit over the five segments yields a drift rate (slope). All per-metric slopes are normalized and aggregated via weighted sum into a single Drift Score. The paper states: "lower is better" and that the metric "summarises overall temporal stability."
  • Inference efficiency: Throughput in frames per second (fps) and per-chunk latency in seconds, measuring wall-clock generation speed.

• Baselines. The paper compares against methods in three categories:

  • Bidirectional diffusion models: LTX-Video (HaCohen et al., 2025)—a real-time Video-VAE with spatiotemporal transformer—and Wan2.1-1.3B (Team, 2025), the foundation model shared by all distilled methods. These are limited to fixed-duration generation (5 seconds for bidirectional models in the evaluation) and cannot scale to arbitrary length.
  • Autoregressive diffusion models: NOVA (Deng et al., 2025)—non-quantized temporal AR with spatial diffusion; Pyramid Flow (Jin et al., 2024b)—pyramidal flow matching with temporal pyramid; SkyReels-V2-1.3B (Chen et al., 2025a)—diffusion forcing with non-decreasing noise schedules; and MAGI-1-4.5B (Teng et al., 2025a)—block-causal attention at 4.5B parameters.
  • Distilled AR models: All sharing the Wan2.1-1.3B backbone and 4-step denoising schedule—CausVid (Yin et al., 2025a)—bidirectional-to-AR DMD distillation; Self-Forcing (Anonymous, 2025)—self-rollout DMD training; and Causal Forcing (Zhu et al., 2025)—diffusion forcing mid-training before self-rollout distillation. The paper states: "All baselines use official checkpoints and are evaluated under identical prompts and generation lengths (5s for bidirectional models)."

• Generation budget / compute accounting. Compute is measured in two ways: throughput and latency for wall-clock efficiency, and equivalent model capacity and denoising steps for quality comparisons. All distilled AR models (CausVid, Self-Forcing, Causal Forcing, HiAR) share the same Wan2.1-1.3B backbone and the same 4-step denoising schedule ($S = 4$), making quality comparisons directly controlled for model capacity and inference compute. HiAR's $\sim 1.8\times$ speedup is measured against this shared baseline: the other distilled models achieve 17 fps and 0.69 s latency, while HiAR achieves 30 fps and 0.30 s latency. The paper notes this speedup "comes at no cost to generation quality; in fact, HiAR simultaneously achieves the best VBench scores and the lowest drift."

• Cross-validation / statistical protocol. The paper does not describe a formal cross-validation protocol for the main experiments. The VBench evaluation uses the standard benchmark protocol on 20-second generations. For ablation studies (Tables 2 and 3), all variants are retrained under matched conditions to ensure train–test consistency: "All variants are retrained under the same rollout mode used at inference to ensure train–test consistency, unless stated otherwise." The drift metric's per-segment linear fits provide some robustness to segment-level variance through the use of slope rather than endpoint differences, but no confidence intervals or statistical significance tests are reported for any of the main results.

Main Quantitative Results

Aggregate Comparison on 20-Second Generation (Table 1)

Table 1 presents the headline comparison across all methods on VBench scores, drift, and inference efficiency at 20 seconds of generation. HiAR achieves the highest Total VBench score (0.821) among all methods—both bidirectional and autoregressive—and the lowest Drift (0.257) among all distilled AR models.

The Quality score is particularly notable: HiAR reaches 0.846, the highest across all methods, surpassing both the bidirectional Wan2.1-1.3B teacher (0.813) and the closest AR competitor Self-Forcing (0.829). This is significant because the Quality sub-score captures visual fidelity dimensions, and HiAR outperforms even the bidirectional model that lacks the error-propagation challenges of autoregressive generation.

On Semantic score, HiAR achieves 0.723, which is lower than Wan2.1-1.3B (0.766) but competitive with or higher than other distilled AR models: Self-Forcing reaches 0.708, Causal Forcing 0.701, and CausVid 0.740. The gap between HiAR and its bidirectional teacher (0.723 vs. 0.766) represents the semantic fidelity cost of switching to few-step autoregressive generation, though HiAR minimizes this cost relative to other AR methods.

The Dynamic score—measuring motion magnitude and variety—tells a particularly important story about the forward-KL regularizer's effectiveness. HiAR achieves 0.686, nearly matching the bidirectional Wan2.1-1.3B teacher (0.690) and substantially outperforming all other AR methods: Self-Forcing drops to 0.542, Causal Forcing reaches 0.672, CausVid is at 0.621. The 0.144 gap between HiAR and Self-Forcing on dynamics (0.686 vs. 0.542) quantifies the motion collapse that the forward-KL regularizer prevents—without it, the model loses over 20% of the teacher's motion diversity.

Drift scores reveal the hierarchical denoising's stability benefits. Among distilled AR models, HiAR achieves the lowest drift at 0.257, compared to 0.355 for Self-Forcing (a 27.6% relative reduction), 0.615 for Causal Forcing, and 0.842 for CausVid. The paper specifically highlights this: "HiAR reduces drift by 27.6% relative to Self-Forcing (0.257 vs. 0.355), confirming that hierarchical denoising with matched context noise levels substantially mitigates the compounding inter-block error that drives long-horizon degradation." CausVid's extremely high drift (0.842) is consistent with the paper's description of "visible colour oversaturation at later segments," while the bidirectional models (LTX-Video, Wan2.1) cannot be evaluated on drift since they are "non-autoregressive and drift is not applicable."

On inference efficiency, the pipelined parallelism yields concrete gains. HiAR achieves 30 fps throughput and 0.30 s per-chunk latency, compared to 17 fps and 0.69 s for all other distilled AR models that share the same Wan2.1-1.3B backbone and 4-step schedule. This is the $\sim 1.8\times$ speedup the paper claims. The larger autoregressive models show substantially worse efficiency: SkyReels-V2-1.3B at 0.49 fps and 112 s latency, MAGI-1-4.5B at 0.19 fps and 282 s latency. The bidirectional models vary widely: LTX-Video achieves 8.98 fps (designed for real-time operation), while Wan2.1-1.3B runs at 0.78 fps.

Qualitative Comparison (Figure 3)

Figure 3 presents visual results across six diverse prompts spanning natural scenery (beach, mountain landscape), objects (umbrellas), and human subjects (rock climbing, woman reading, baby portrait). The paper describes a consistent degradation hierarchy across the baselines:

CausVid "exhibits the most severe degradation: frames progressively shift toward neon green and yellow tints, with scene content largely unrecognisable by 20 s." This extreme color drift is consistent with its 0.842 Drift score—the highest in Table 1.

Self-Forcing and Causal Forcing "alleviate this to some extent, yet still develop visible colour oversaturation and hue drift over time." The paper notes degradation is "particularly pronounced on human-centric content—facial regions suffer from unnatural colour casts and loss of fine detail (e.g., skin texture, facial features), which are perceptually salient and difficult to mask." This observation about human-face sensitivity is practically important: perceptual metrics may weight facial quality heavily, and face-specific degradation might be a particularly damaging failure mode for applications involving human subjects.

HiAR "maintains stable colour fidelity, sharpness, and structural coherence from the first frame to the last across all content types, with no perceptible drift in either scenery or portrait prompts." This visual assessment aligns with the quantitative drift score of 0.257.

The paper does not provide quantitative per-prompt breakdowns or user study results, so the qualitative comparison relies on visual inspection of the sampled frames in Figure 3 rather than statistical analysis.

Context Noise Level Ablation (Table 2)

Table 2 reports the effect of three context noise level settings on Quality, Semantic, Smooth (VBench motion smoothness), and Drift. All variants are retrained under their respective noise level configurations to ensure train–test consistency. The three settings map directly to the theoretical analysis in Section 3.1:

1. $t_c = t_j$ (input noise level): The context is presented at the same noise level as the current block's input, meaning block $B_n$ cannot observe the result of denoising step $j$ on block $B_{n-1}$. This "removes intra-step causality" as the paper states. The result: lowest Drift (0.184) but substantially degraded quality—Quality drops to 0.799 (from 0.846 for the default) and Semantic drops to 0.692 (from 0.723). Motion smoothness also suffers at 0.978, the lowest of the three settings. This configuration achieves the best error attenuation (most noise diluting the bias) but at the cost of breaking temporal coherence because the model cannot see what happened to the previous block during the current denoising step.

2. $t_c = t_{j+1}$ (output noise level; HiAR default): The "matched-noise" setting derived from the optimality analysis in Section 3.1. Results: Quality 0.846, Semantic 0.723, Smooth 0.988, Drift 0.257. This achieves the best Quality and Semantic scores among the three settings, with motion smoothness (0.988) only slightly below the Self-Forcing setting. The drift reduction relative to $t_c = 0$ (0.257 vs. 0.355) is substantial, confirming the theoretical prediction that matched-noise context attenuates error propagation while preserving sufficient information for temporal coherence.

3. $t_c = 0$ (clean context; standard Self-Forcing): The conventional approach of providing fully denoised predictions as context. Results: Quality 0.829, Semantic 0.708, Smooth 0.991, Drift 0.355. This achieves the best motion smoothness (0.991) but the highest drift (0.355), consistent with the paper's core argument that clean context maximizes error propagation. Quality and Semantic are both lower than the $t_{j+1}$ setting, indicating that the error accumulation from clean context degrades not just temporal stability but also average visual quality across the full 20-second generation.

The paper summarizes: the $t_{j+1}$ setting "strikes the optimal balance: it preserves nearly the same temporal smoothness as Self-Forcing (0.988 vs. 0.991) while substantially reducing drift and improving overall quality." The smoothness gap (0.988 vs. 0.991) is minimal, suggesting that matched-noise context does not meaningfully compromise frame-to-frame coherence despite providing noisier conditioning information.

Forward-KL Regularizer Design Ablation (Table 3)

Table 3 systematically ablates the design choices of the forward-KL regularizer, reporting Quality, Semantic, Dynamic, and Drift. This is where the paper's motion diversity claims are empirically grounded.

Attention mode comparison.

• Bi-attn + 1 step (default): Dynamic 0.686, Quality 0.846, Semantic 0.723, Drift 0.257.
• Causal + 1 step: Dynamic drops to 0.625, Quality drops to 0.828, Semantic drops to 0.701, Drift rises slightly to 0.271.

The paper explains this performance gap through the qualitative difference in how the two attention modes produce frames (illustrated in Figure 5). Under bidirectional attention, all frames at a given denoising step exhibit "uniform quality and blur." Under causal attention, frames become "progressively sharper along the temporal axis" because preceding frames fix low-frequency structure and reduce uncertainty for later frames. The teacher's bidirectional trajectory targets—which have uniform refinement patterns—are therefore "mismatched with the student's causal-mode outputs," making direct causal-mode forward-KL less effective. The bidirectional-attention decoupling strategy exploits the empirical correlation shown in Figure 4 (Pearson $r = 0.968$) to regularize dynamics through the well-matched bidirectional pathway without interfering with the causal DMD loss.

Number of constrained steps ($K$).

• $K = 1$ (default): Dynamic 0.686, Quality 0.846, Semantic 0.723, Drift 0.257.
• $K = 2$: Dynamic rises slightly to 0.693, but Quality drops to 0.835 and Drift worsens to 0.296.
• $K = 4$: Dynamic is essentially flat at 0.691, while Quality drops further to 0.813 and Drift to 0.306.

The marginal dynamic gains from constraining more steps are minimal ($< 0.01$ beyond $K = 1$), while the quality degradation is monotonic and substantial (0.846 → 0.835 → 0.813). This supports the paper's claim that motion diversity is "primarily governed by the low-frequency structure laid down in the first denoising step." Constraining later steps provides "diminishing returns while interfering with the model's denoising capacity"—the high-frequency refinement in later steps benefits from the student learning its own distribution rather than being forced to match specific teacher trajectories. This is an interesting finding about the phase structure of video diffusion: optimization pressure that helps in early stages (where global structure is determined) hurts in later stages (where detail generation requires flexibility).

Component necessity ablations.

• w/o LFKL (no forward-KL regularizer): Dynamic collapses to 0.445—a 35% drop from the full model—while Quality (0.839), Semantic (0.732), and Drift (0.218) remain competitive or even improve. This is the key evidence for the "low-motion shortcut" diagnosis: the model can achieve good visual quality and temporal stability without forward-KL, but at the cost of producing near-static videos. The paper's statement that "the model falls into the low-motion shortcut without forward-KL regulation" is directly supported: the 0.445 Dynamic score is the lowest across all ablations and baselines in both Table 1 and Table 3.

• w/o re-training (hierarchical denoising at inference only): Quality plummets to 0.767 (from 0.846), Semantic drops to 0.559 (from 0.723), while Dynamic falls to 0.512 and Drift rises to 0.309. This demonstrates the train–test mismatch problem: applying the step-first generation order to a model trained with block-first order produces severe quality degradation. The paper notes this "significantly reduced drift compared with Self-Forcing (0.309 vs. 0.355) yet at a substantial cost to visual quality (Quality 0.767), highlighting the importance of train–test alignment."

• w/o hierarchical denoising (Self-Forcing baseline): This recovers the standard Self-Forcing configuration: Quality 0.829, Semantic 0.708, Dynamic 0.542, Drift 0.355. Compared to the full HiAR, Quality is lower (0.829 vs. 0.846), Dynamic is substantially lower (0.542 vs. 0.686), and Drift is higher (0.355 vs. 0.257). This triple disadvantage validates the paper's claim that hierarchical denoising and forward-KL regularization each contribute independently to the overall improvement.

A subtle pattern across the ablations: there appears to be a quality–drift trade-off in some configurations. The "w/o LFKL" variant achieves the lowest Drift (0.218) but at the cost of collapsed dynamics (0.445). The "$t_c = t_j$" variant achieves very low Drift (0.184) but poor Quality (0.799). The full HiAR with $K = 1$ achieves Quality 0.846 and Drift 0.257—not the absolute best on either dimension individually, but the best balanced configuration. This suggests that temporal stability alone is not a sufficient objective; the model can achieve stability through degenerate solutions (static output, no intra-step causality), and the optimal configuration requires explicit regularization to maintain both stability and quality.

Correlation Analysis Between Attention Modes (Figure 4)

Figure 4 provides empirical justification for the bidirectional-attention decoupling strategy by tracking Dynamic scores under both attention modes across training checkpoints without forward-KL regularization. Each point represents one training checkpoint; the x-axis is the Bidirectional Dynamic Score, the y-axis is the Causal Dynamic Score, and color encodes training step (from 0 to 20,000 steps).

The key result is a strong positive correlation: Pearson $r = 0.968$ with $p < 10^{-6}$. The paper interprets this as confirming that "the low-motion shortcut affects both attention modes simultaneously and that regularising the bidirectional mode effectively constrains causal-mode dynamics."

The training-step color encoding reveals an additional dynamic: both scores decline consistently as training progresses (the points move from upper-right to lower-left over the 20,000 training steps), confirming that the low-motion collapse is a progressive training phenomenon, not a sudden collapse or a static property. Without regularization, both attention modes converge toward lower Dynamic scores together, and the strong correlation means that any intervention arresting the bidirectional decline will likely arrest the causal decline as well.

The confidence interval (95% CI) is shown on the plot and is tight around the regression line, indicating the correlation is reliable across training checkpoints rather than driven by a few outliers. The paper does not report the specific $R^2$ value or the regression equation, but the Pearson $r$ of 0.968 implies $R^2 \approx 0.937$—bidirectional dynamics explain approximately 93.7% of the variance in causal dynamics during training.

Qualitative Attention Mode Comparison (Figure 5)

Figure 5 visualizes the difference between bidirectional and causal attention in single-step denoising outputs, providing intuition for why bidirectional-mode forward-KL targets are better matched than causal-mode targets. Under bidirectional attention, frames across all temporal positions (0s, 1s, 2s, 5s) exhibit "uniform quality and blur"—every frame is at the same stage of refinement because full-sequence attention treats all positions symmetrically. Under causal attention, frames show progressive sharpening: earlier frames appear more refined while later frames retain more blur, because "as preceding frames fix the low-frequency structure, the conditional distribution of later frames concentrates, resulting in higher-frequency details."

This asymmetry has direct implications for the forward-KL loss. The teacher's bidirectional trajectory produces frames with uniform refinement levels at each step, which matches the student's bidirectional outputs well. But when the student operates in causal mode, the frame-by-frame refinement level varies systematically with temporal position, creating a mismatch between the teacher's uniform targets and the student's position-dependent outputs. The paper states: "a distillation target derived from bidirectional denoising provides a spatiotemporally uniform supervision signal well suited to regularising global dynamics, whereas directly constraining causal outputs introduces mismatched targets that are tightly coupled with the model's autoregressive generation pathway, degrading overall quality."

Ablation Studies and Robustness Checks

Context noise level ($t_c$): The three-point comparison in Table 2 tests the extremes of the feasible region identified in Section 3.1. The $t_c = t_{j+1}$ setting (matched-noise) achieves the best Quality (0.846) and Semantic (0.723) while maintaining high Smooth (0.988) and reducing Drift by 27.6% relative to $t_c = 0$ (0.257 vs. 0.355). The non-obvious finding is that $t_c = t_j$ (input level, noisier than the optimal) achieves the lowest Drift (0.184) but substantially degrades Quality (0.799) and Smooth (0.978)—error attenuation can be pushed too far, breaking temporal continuity in exchange for stability.

Forward-KL attention mode: The bidirectional vs. causal comparison in Table 3 reveals that applying forward-KL in causal mode degrades both dynamics (0.625 vs. 0.686) and quality (0.828 vs. 0.846). The Pearson correlation in Figure 4 ($r = 0.968$) provides the empirical justification for the decoupling strategy, but the ablation reveals that correlation alone is insufficient—the causal-mode forward-KL still underperforms despite the correlation, indicating that direct causal constraint introduces gradient interference absent in the bidirectional-only approach.

Forward-KL step count ($K$): The sweep from $K = 1$ to $K = 4$ in Table 3 reveals diminishing and then negative returns: Dynamic improves marginally (0.686 → 0.693 → 0.691), but Quality degrades monotonically (0.846 → 0.835 → 0.813) and Drift worsens (0.257 → 0.296 → 0.306). This is a clear "sweet spot" finding: $K = 1$ is optimal, and more is worse. The paper's interpretation—that early steps govern motion layout while later steps handle detail refinement that benefits from flexibility—is supported but not directly tested (there is no experiment that isolates low-frequency from high-frequency contributions to dynamics).

Component necessity: The "w/o LFKL" ablation in Table 3 is the most striking negative result: removing the forward-KL regularizer causes Dynamic to collapse to 0.445, confirming that the low-motion shortcut is real and severe. The "w/o re-training" ablation demonstrates that hierarchical denoising cannot be applied as an inference-time modification to existing models—Quality drops to 0.767 and Semantic to 0.559, confirming the train–test mismatch hypothesis. The "w/o hierarchical denoising" row (Self-Forcing baseline) shows the triple penalty of the conventional approach: lower Quality (0.829), lower Dynamic (0.542), and higher Drift (0.355).

Distilled model parity: All distilled models (CausVid, Self-Forcing, Causal Forcing, HiAR) share the same Wan2.1-1.3B backbone and 4-step schedule, so architecture and compute are controlled. Performance differences can therefore be attributed to the generation/training methodology rather than model capacity or inference budget. This is a strength of the experimental design—but it also means the results are conditional on the specific Wan2.1 architecture and may not transfer to other backbones.

Long-horizon consistency: The drift metric's five-segment design ensures that the 20-second evaluation actually captures temporal degradation rather than averaging it away. The qualitative results in Figure 3 visually corroborate the drift rankings: CausVid's 0.842 corresponds to extreme color shifts, while HiAR's 0.257 corresponds to stable output. The paper does not report per-segment VBench scores or per-segment drift component breakdowns, which would provide a more granular view of how degradation progresses over the 20-second horizon.

Generalization beyond training duration: The paper trains on 5-second clips but evaluates at 20 seconds (4× the training duration). The strong performance at 20 seconds indicates that the hierarchical denoising and forward-KL regularizer produce models that generalize beyond their training horizon—an important robustness property that is demonstrated but not explicitly discussed or ablated. There is no experiment testing at even longer durations (e.g., 40 or 60 seconds) to determine where the generalization breaks down.

Critical Assessment

Claim 1: Hierarchical denoising with matched-noise context reduces drift by suppressing error propagation.

What the experiments demonstrate: Table 1 shows HiAR achieves Drift 0.257 vs. Self-Forcing's 0.355—a 27.6% reduction. Table 2 shows that $t_c = t_{j+1}$ (matched-noise) achieves Drift 0.257 vs. $t_c = 0$ (clean context) at 0.355, directly linking the noise level to the drift reduction. The qualitative results in Figure 3 visually corroborate the stability improvement.

What the experiments do not demonstrate: The error decomposition in Section 3.1 predicts that matched-noise context reduces error propagation by the attenuation coefficient $(1 - \sigma_{t_c})$, but the paper never measures the actual prediction error $\delta^{(n)}$ or validates that the error propagation follows the predicted pattern. The drift metric measures perceptual degradation, not the inter-block prediction error that the theoretical analysis is about. A direct experiment would inject known errors into context at different noise levels and measure how much they affect subsequent blocks—this is absent. The claim that drift reduction is "caused by" error attenuation via matched-noise context is therefore supported by the theoretical framework and the empirical correlation between noise level and drift, but the causal mechanism is inferred rather than directly validated.

Additionally, the drift metric is introduced in this paper and has no established correlation with human judgments of temporal quality. While the components (MUSIQ, DINOv2 similarity, LPIPS) are individually validated in prior work, their combination and the linear-fit-over-segments aggregation is novel and unvalidated. The paper does not report a human evaluation or correlation study establishing that lower Drift scores correspond to perceptually preferred videos.

Claim 2: HiAR achieves the best overall VBench score (0.821) and the lowest drift (0.257) among all compared methods.

What the experiments demonstrate: Table 1 reports these exact numbers. HiAR's Total score (0.821) exceeds Wan2.1-1.3B (0.802), Self-Forcing (0.805), Causal Forcing (0.810), and all other methods. The Drift score (0.257) is the lowest among all AR models for which drift is applicable.

Where the evidence is weaker: The VBench evaluation is conducted on 20-second generations for AR models but only 5-second generations for bidirectional models. This is an apples-to-oranges comparison: bidirectional models are evaluated within their training distribution (short clips), while AR models are tested on a 4× extrapolation beyond their training clip duration (5 s → 20 s). HiAR's Total score exceeding Wan2.1-1.3B (0.821 vs. 0.802) is impressive, but the bidirectional model might perform differently (potentially worse due to fixed-window limitations, or potentially better if evaluated on its native duration) if it could be evaluated on 20-second outputs. The paper acknowledges this asymmetry ("5s for bidirectional models") but does not discuss its implications for the fairness of the comparison.

Furthermore, the number of evaluation prompts is not specified. VBench is a standard benchmark, but the paper doesn't report the sample size (number of generated videos) used for the evaluation. Without this, the statistical reliability of the differences—particularly the small gaps between top methods (0.821 vs. 0.810 vs. 0.805)—cannot be assessed.

Claim 3: The forward-KL regularizer in bidirectional-attention mode prevents motion collapse, preserving dynamics near the teacher's level (0.686 vs. 0.690).

What the experiments demonstrate: Table 3 shows Dynamic collapses from 0.686 to 0.445 when LFKL is removed, directly supporting the claim. Figure 4 shows the strong correlation ($r = 0.968$) between bidirectional and causal dynamics, justifying the decoupling strategy. Table 3 also shows that applying LFKL in causal mode is less effective (Dynamic 0.625), validating the bidirectional-mode design choice.

What the experiments do not demonstrate: The forward-KL loss (Equation 10) is motivated as approximating a forward-KL divergence that encourages mode-covering, but the paper provides no evidence that the trained model actually covers more of the teacher's output modes—only that the VBench Dynamic score (which measures aggregate motion magnitude and variety) is higher. Mode collapse and motion diversity are related but distinct concepts: a model could produce diverse motion patterns that all cluster in a single "high-motion" mode of the distribution, or a model could cover multiple modes but with reduced motion within each. The paper does not present any distribution-level analysis (e.g., comparing the diversity of generated motion vectors, or measuring coverage of the teacher's motion distribution) that would directly validate the mode-covering mechanism.

Additionally, the claim that hierarchical denoising "amplifies the low-motion shortcut" is stated as the motivation for needing forward-KL, but there is no experiment comparing the motion collapse rate with and without hierarchical denoising under the same reverse-KL objective. The paper compares Self-Forcing (block-first, Dynamic 0.542) to HiAR without LFKL (step-first, Dynamic 0.445), but these differ in both generation order and the presence of forward-KL. To isolate the amplification effect, one would need to compare Dynamic scores after matched training steps for block-first vs. step-first training without forward-KL—this ablation is not reported.

Claim 4: Pipelined parallelism yields $\sim 1.8\times$ wall-clock speedup (30 fps vs. 17 fps).

What the experiments demonstrate: Table 1 reports these throughput numbers, and the paper provides a detailed description of the pipelining and KV-cache fusion optimizations. The speedup is measured against other distilled AR models with the same backbone and denoising steps.

Where the evidence is weaker: The speedup is reported for a specific configuration (4 denoising steps, $N$ blocks, specific hardware). The paper does not report how the speedup scales with the number of denoising steps—a 4-step schedule is unusually aggressive (most diffusion models use 20–50 steps). If denoising steps were increased (e.g., for higher-quality generation), the relative benefit of pipelined parallelism might change (more steps → more pipeline stages → potentially higher utilization, but also more communication overhead). The paper does not discuss the hardware used for timing measurements or whether the comparison controls for implementation differences (e.g., the other models might not implement KV-cache fusion, which is an optimization orthogonal to hierarchical denoising).

Claim 5: Retraining under hierarchical denoising is necessary—inference-only application causes severe quality degradation.

What the experiments demonstrate: Table 3 ("w/o re-training"): Quality drops from 0.846 to 0.767, Semantic from 0.723 to 0.559. This is a clear and large effect.

What remains unclear: The "w/o re-training" ablation applies hierarchical denoising at inference to a model that was trained with Self-Forcing (block-first, clean context). But what if the model were trained with diffusion forcing (varying noise levels, but still block-first) instead? Would the partial robustness to noisy contexts from diffusion forcing training partially close the gap, making retraining less critical? The paper doesn't test this intermediate case. The "w/o re-training" result establishes that naive application fails, but doesn't map the boundary of where train–test mismatch becomes severe enough to require full retraining.

Missing experiments that would strengthen the paper:

1. Per-segment drift decomposition: Reporting individual drift components (MUSIQ slope, DINOv2 similarity slope, saturation slope, etc.) would reveal which degradation types are most and least improved by hierarchical denoising. Does HiAR prevent color drift specifically, or motion repetition, or both?

2. Longer-duration evaluation: The paper evaluates at 20 seconds (4× training clip duration). Testing at 40, 60, or 120 seconds would reveal whether the drift reduction is sustained or whether HiAR eventually succumbs to degradation at longer horizons, and would test the limits of the sliding-window KV cache design.

3. Direct error propagation measurement: Injecting synthetic errors of known magnitude into context at different noise levels and measuring the error in subsequent blocks would provide direct causal evidence for the bias–information trade-off mechanism.

4. Human evaluation: Given that video quality is ultimately a perceptual judgment, a human study comparing HiAR against Self-Forcing and Causal Forcing on temporal consistency, motion naturalness, and overall quality would substantially strengthen the claims, particularly for the drift metric which is novel and unvalidated.

5. Backbone ablation: All experiments use Wan2.1-1.3B. Testing hierarchical denoising on a different base architecture (e.g., a different DiT variant, or a UNet-based video diffusion model) would establish whether the matched-noise principle and reordering benefits are architecture-specific or general.

6. Step-count scaling: The 4-step schedule is fixed throughout. Ablating the number of denoising steps (e.g., 2, 8, 16 steps) would reveal how the benefits of hierarchical denoising scale with inference compute—does the drift reduction become more or less pronounced with more steps?

7. Training duration ablation: The paper trains for 20,000 steps. Does the low-motion shortcut worsen with more training? Does forward-KL prevent collapse indefinitely, or does it only delay it? A longer training run with and without LFKL would characterize the training dynamics more completely.

8. Combination with other anti-drift methods: HiAR addresses drift through context noise levels. Other methods address drift through different mechanisms (e.g., Bagger's backwards aggregation, Po et al., 2025; diffusion forcing's noisy-context training). Are these complementary? Could hierarchical denoising be combined with diffusion forcing training to further reduce drift?

6. Limitations and Trade-offs

6.1 Single Backbone and Single Benchmark: Generality Is Unproven

The assumption or constraint. All experiments—both the main results (Table 1) and every ablation (Tables 2, 3)—use a single model backbone (Wan2.1-1.3B) and a single evaluation benchmark (VBench). The paper states this explicitly in Section 4.1: "We use the Wan2.1-1.3B backbone as our base model." There is no experiment with a different DiT architecture, a UNet-based video diffusion model, or even a different parameter scale within the Wan family. The evaluation is conducted entirely on VBench; no results are reported on other video generation benchmarks (e.g., UCF-101, Kinetics, MSR-VTT, or any of the VBench-2.0 extensions that the paper cites in its references).

The consequence. Without multi-backbone or multi-benchmark evidence, three important uncertainties remain unresolved:

First, the matched-noise optimality principle ($t_c^* = t_{j+1}$, derived in Section 3.1) is architecture-agnostic in its derivation, but its empirical validation may depend on properties of the Wan2.1 backbone—its noise schedule (the shift parameter $s$ in Equation 1), its attention mechanism, its pretraining data distribution. A different architecture with a different noise schedule curvature might shift the bias–information trade-off such that the boundary optimum $t_c = t_{j+1}$ is no longer strictly best, or such that the quality penalty for $t_c = t_j$ (Table 2, Quality 0.799) becomes less severe.

Second, the low-motion shortcut and the effectiveness of the forward-KL regularizer may depend on the specific distillation setup. The paper identifies the mode-seeking tendency of reverse-KL DMD as the root cause (Section 3.3), but the rate and severity of collapse likely depends on the teacher–student capacity gap (Wan2.1-14B teacher vs. 1.3B student), the number of denoising steps ($S = 4$), and the training data distribution. A different teacher–student pair or a different step count might exhibit different collapse dynamics.

Third, VBench-specific behaviors cannot be ruled out. VBench's 16 dimensions and their aggregation weights may favor particular types of motion, quality, or semantic alignment that interact with HiAR's design in uncharacterized ways. The Dynamic score, for instance, measures motion magnitude and variety through specific metrics that might not capture all forms of motion diversity collapse. A model that performs well on VBench dynamics might still exhibit perceptually unnatural motion patterns that a different benchmark would detect.

What evidence exists in the paper. None. The paper contains no cross-backbone or cross-benchmark experiments. The only model variation is the teacher model size (1.3B vs. 14B) for the DMD critic, which tests critic capacity, not generator architecture. The paper does not discuss this limitation or claim generality beyond the tested configuration.

Mitigation status. Not addressed. The paper does not acknowledge the single-backbone limitation or suggest that future work should validate the hierarchical denoising principle on other architectures. The absence of this caveat is a notable omission given that the paper's central theoretical claim—the matched-noise optimality—is presented as a general principle (Section 3.1) without architecture-specific qualifiers.

---

6.2 Training Cost and Complexity: A Full Retraining Requirement

The assumption or constraint. HiAR requires training from scratch (or at minimum, full fine-tuning of a pretrained model) under the hierarchical denoising schedule. The paper demonstrates this requirement empirically in Table 3 ("w/o re-training"): applying hierarchical denoising at inference only—without retraining—causes Quality to drop from 0.846 to 0.767 and Semantic to drop from 0.723 to 0.559. The paper states this is necessary because "a train–test gap remains when the model has been trained under the conventional block-first rollout" (Section 3.3).

The training procedure itself is complex and expensive relative to standard teacher-forced training. It involves: (1) sampling 16,000 ODE solution pairs from the base model for DMD distillation; (2) sampling 20,000 additional 50-step ODE trajectories for forward-KL regularization; (3) training a separate critic model (Wan2.1-14B) adversarially against the student at a 5:1 update ratio; (4) performing self-rollout training where the student generates its own context during each training iteration; (5) running forward-KL trajectory matching in bidirectional-attention mode in parallel with the causal DMD loss. The total training budget is 20,000 steps at batch size 64 on 5-second clips.

The consequence. The practical barrier to adopting HiAR is substantial. A practitioner with an existing pretrained bidirectional video diffusion model cannot simply apply HiAR as an inference-time improvement—they must undertake a full distillation and retraining pipeline. This is fundamentally different from methods that offer training-free or lightweight-finetuning solutions. Specifically:

The compute cost of the training pipeline is dominated by three factors: the 20,000 teacher trajectory samples (each requiring 50 full ODE integrations of the base model), the adversarial critic training (running Wan2.1-14B, a model with roughly 10× the parameters of the student), and the self-rollout procedure (which requires generating context during training rather than reading ground-truth context from disk). The paper does not report total GPU-hours or compare training cost to Self-Forcing or other baselines, making it impossible for practitioners to estimate the resource requirements.

The data requirements are model-specific. The 16,000 ODE pairs and 20,000 trajectories must be sampled from the same base model being distilled, because the teacher trajectories define the target distribution. If a practitioner wanted to apply HiAR to a different base model, they would need to regenerate all training data from scratch using that model. This is not a one-time cost amortized across users; it is a per-model cost.

The training instability risk is real. The paper documents that without forward-KL regularization, dynamics collapse to 0.445 (Table 3, "w/o LFKL"), and that the ReST$^{EM}$-style training attempted in a related context (Appendix K, though this is referenced for the revision model in the prior work discussion) can backfire. The training recipe—DMD with critic at 5:1 ratio, forward-KL with $\lambda = 0.1$, $K = 1$, bidirectional-only mode—is the product of extensive empirical tuning, and there is no guarantee that these hyperparameters transfer to a new base model or different video domain without similar tuning effort.

What evidence exists in the paper. Table 3 ("w/o re-training") provides direct evidence that retraining is necessary for the Wan2.1-1.3B backbone with the specified training configuration. No experiments test whether partial retraining (e.g., fine-tuning only the last few layers, or training for fewer steps) could recover some of the quality loss. No experiments compare the total training FLOPs of HiAR versus Self-Forcing or Causal Forcing to quantify any training-time premium for the hierarchical schedule.

Mitigation status. The paper acknowledges the retraining requirement implicitly (Section 3.3: "We therefore retrain with self-rollout under the hierarchical schedule") but does not frame it as a limitation, discuss its practical cost, or compare training efficiency with alternatives. The paper does not suggest directions for reducing the retraining burden (e.g., whether diffusion forcing pretraining could serve as a better initialization, reducing the number of hierarchical training steps needed).

---

6.3 Evaluation Duration Cap: No Evidence Beyond 20 Seconds

The assumption or constraint. All quantitative evaluations in the paper are conducted on 20-second video generations. This applies to the VBench scores in Table 1, the Drift metric in Table 1, the ablations in Tables 2 and 3, and the qualitative comparisons in Figure 3. The training is performed on 5-second clips, so the evaluation already represents a 4× extrapolation beyond the training duration. The paper presents this as a strength: HiAR maintains quality at 4× the training horizon without explicit long-duration training.

The paper's motivation, however, frames the problem as "indefinite extension" and "theoretically infinite length" (Abstract, Section 1), and cites applications in "interactive agents and world models" (Section 1) that require continuous, sustained video output. A 20-second evaluation, while longer than typical for video generation benchmarks, does not test indefinite-length generation in any meaningful sense.

The consequence. Without longer-duration evaluation, several questions central to the paper's claims remain unanswered:

Does drift eventually accumulate beyond 20 seconds? The drift metric in Table 1 measures the linear slope of degradation over five 4-second segments. A 27.6% reduction in drift relative to Self-Forcing (0.257 vs. 0.355) is demonstrated at 20 seconds, but this does not guarantee the reduction is sustained at 60 seconds, 5 minutes, or longer. The sliding-window KV cache (5-second attention window, Section 4.1) means the model has no access to frames beyond the most recent 5 seconds. If error accumulation has a nonlinear component—if errors compound multiplicatively rather than additively—the 20-second evaluation window might be too short to observe the regime where degradation accelerates. The paper's error decomposition in Section 3.1 predicts additive error propagation (the bias term $(1 - \sigma_{t_c}) \delta^{(n-1)}$ adds linearly at each step), but this is a local, single-step analysis that does not account for the model's response to accumulated errors over many blocks.

Does motion diversity remain stable over long horizons? The Dynamic score in Table 1 is an aggregate measure over the full 20-second video. It does not reveal whether motion diversity changes over time—for example, whether the model gradually shifts from diverse motion in early segments to repetitive or diminished motion in later segments. The forward-KL regularizer is designed to prevent global motion collapse during training, but it does not explicitly target temporal consistency of motion diversity during long rollouts. A model could maintain good aggregate Dynamic scores while becoming progressively more static over the course of a long generation.

Does the matched-noise principle break down at very long horizons? The optimality derivation in Section 3.1 assumes that the prediction error $\delta^{(n-1)}$ has finite variance and that the model's response to noisy context is well-approximated by the local linear analysis. At very long horizons, the accumulated error might grow to the point where the matched-noise attenuation is insufficient—the bias term, even attenuated by $(1 - \sigma_{t_c})$, might become large enough to push the context out of the distribution the model was trained to handle. The training data consists entirely of 5-second clips with self-rollout errors accumulated over at most a few blocks; the model has never seen the large accumulated errors that would arise after hundreds of autoregressive steps.

What evidence exists in the paper. None. There are no experiments at 40, 60, 120 seconds, or longer. The paper does not report any analysis of how drift evolves beyond the 20-second window, nor does it characterize the empirical scaling of drift with generation duration. The qualitative results in Figure 3 show 20 seconds of output, and within that window HiAR appears stable, but the paper does not claim or demonstrate stability beyond this.

Mitigation status. Not addressed. The paper does not acknowledge the gap between the 20-second evaluation and the "indefinite extension" motivation. There is no discussion of what limits the maximum generation duration, what failure modes might emerge at longer horizons, or what future work would be needed to characterize or extend the stable generation horizon.

---

6.4 Drift Metric Is Unvalidated and Its Components Are Underexplained

The assumption or constraint. The paper introduces a custom "drift metric suite" as a key evaluation tool. The metric is described in Section 4.1: each 20-second video is divided into five equal temporal segments, per-segment statistics are computed across multiple dimensions (MUSIQ, CLIP-IQA for perceptual quality; DINOv2 consecutive-frame cosine similarity and LPIPS consecutive-frame distance for temporal coherence; HSV saturation mean and Laplacian variance for low-level statistics), a linear fit over the five segments yields a per-metric drift rate, and these rates are normalized and aggregated via weighted sum into a single Drift Score. The paper states that the score "summarises overall temporal stability."

The metric serves as the quantitative foundation for the paper's central claim about drift reduction. The Abstract highlights that HiAR achieves "the lowest temporal drift among all compared methods." The 27.6% drift reduction relative to Self-Forcing (0.257 vs. 0.355) is one of the paper's headline numbers.

The consequence. A custom metric that lacks validation against human judgments or established evaluation protocols creates several interpretive problems:

Unknown correlation with perceptual quality. The weighted sum of normalized slopes across six disparate statistics (MUSIQ, CLIP-IQA, DINOv2 similarity, LPIPS distance, HSV saturation, Laplacian variance) produces a single number whose relationship to human perception of temporal degradation is entirely unknown. A video with a drift score of 0.257 might be perceptually indistinguishable from one with 0.355, or the difference might be stark. The paper provides no human evaluation—no user study, no Mean Opinion Score (MOS) data, no correlation analysis between Drift Scores and human judgments of temporal consistency. The qualitative results in Figure 3 visually corroborate the drift rankings (CausVid at 0.842 looks worse than HiAR at 0.257), but this only validates the metric at extremes; it does not establish that the metric is sensitive or well-calibrated in the region where methods differ by smaller margins (e.g., HiAR at 0.257 vs. Self-Forcing at 0.355).

Undisclosed aggregation weights. The paper states that per-metric slopes are "normalised and aggregated via a weighted sum" but does not report the normalization procedure or the weights. Different normalization choices (z-score normalization vs. min-max scaling vs. division by standard deviation) produce different relative contributions from each component metric. Different weighting schemes (uniform vs. emphasizing perceptual quality vs. emphasizing temporal coherence) would change the relative rankings of methods. Without this information, the Drift Score is not reproducible—a different lab measuring the same videos with the same per-segment statistics but different normalization/weighting choices could obtain different drift rankings.

Component-level results are absent. The paper reports only the aggregated Drift Score. It does not report per-component drift rates, so it is impossible to determine which types of degradation HiAR prevents most effectively. Does HiAR primarily reduce color drift (HSV saturation slope), motion repetition drift (DINOv2 similarity slope), perceptual quality drift (MUSIQ slope), or all equally? Different applications might care about different degradation types—a world model application might prioritize motion consistency over color fidelity, for instance—and without component-level results, practitioners cannot assess whether HiAR's drift reduction matches their specific requirements.

The linear-fit assumption may be inappropriate. The drift metric assumes that degradation is approximately linear over the five segments (otherwise the slope of a linear fit is not a meaningful summary). The paper provides no evidence for this assumption. If degradation is nonlinear—for example, remaining stable for the first 15 seconds and then degrading rapidly—the linear slope would misrepresent the temporal quality profile. The paper does not report $R^2$ values for the linear fits, plot the per-segment statistics to allow visual inspection of linearity, or compare linear vs. nonlinear trend models.

What evidence exists in the paper. The paper reports Drift Scores for all compared AR methods in Table 1 and for all ablation variants in Tables 2 and 3. The qualitative results in Figure 3 provide visual evidence that high drift scores (CausVid at 0.842) correspond to visible degradation and low drift scores (HiAR at 0.257) correspond to stable output, but this validates only the metric's ability to distinguish extreme cases. No human evaluation, component-level breakdown, or linearity analysis is provided.

Mitigation status. Not addressed. The paper does not discuss the metric's limitations, validate it against human judgment, report component-level results, or provide the normalization/weighting details needed for reproduction. The metric is treated as a reliable evaluation instrument, but its reliability is assumed rather than established.

---

6.5 The Generalization Gap from 5-Second Training to Longer Generation Is Characterized Only at a Single Extrapolation Point

The assumption or constraint. All training is conducted on 5-second video clips (Section 4.1: "We train... for 20k steps on 5-second clips"). The generation evaluation is performed at 20 seconds—a 4× extrapolation. The model uses a sliding-window KV cache with a 5-second attention window at inference (Section 4.1: "we employ a sliding-window KV cache with a constant attention window of 5 s"). This means the model has no access to any frame more than 5 seconds in the past when generating later blocks.

The paper presents the strong performance at 20 seconds as a success: HiAR generalizes beyond its training horizon without explicit long-duration training. This is indeed a positive result, but the paper's framing in Section 1—"generating videos of theoretically infinite length" and "indefinite extension"—implies that the 4× extrapolation is just the first step toward much longer generation. The training design (5-second clips, 5-second attention window) creates a specific generalization challenge that is tested at only a single extrapolation factor.

The consequence. The single-extrapolation-point evaluation leaves fundamental questions about the generalization behavior unresolved:

How does performance scale with generation duration? The paper provides one data point at 20 seconds (4× training duration) with no measurements at intermediate (10 seconds, 15 seconds) or longer (40, 60, 120 seconds) durations. We do not know whether the drift rate is constant (drift accumulates linearly with duration), accelerating (errors compound, drift accelerates), or decelerating (the model stabilizes after an initial transient). The linear slope from the drift metric (fitted over five segments within 20 seconds) cannot answer this question—it characterizes only the rate of change within the 20-second window, not how that rate might change at longer horizons.

Does the 5-second attention window create a hard information bottleneck? With only 5 seconds of temporal context, the model cannot condition on events or motions that occurred more than 5 seconds ago. For short clips (5 seconds is the training duration), this is not a limitation—the entire video fits within the window. At 20 seconds, the model must generate coherent continuations without access to the video's beginning, relying only on the most recent 5 seconds to maintain narrative, stylistic, and motion consistency. At 60 seconds or 5 minutes, this information bottleneck becomes increasingly severe. The paper demonstrates that 5 seconds of context is sufficient for 20 seconds of coherent output, but whether it is sufficient for 60 seconds or indefinite generation is unknown and untested.

Does the model develop long-horizon memory implicitly? Some autoregressive models develop implicit memory mechanisms where information propagates forward through the sequence even without explicit long-range attention—the hidden state of the generation process might encode information about earlier frames that is not captured by the explicit KV cache. The paper does not investigate whether HiAR exhibits such behavior, at what rate long-horizon information decays, or whether information beyond the 5-second window is preserved in any form.

What evidence exists in the paper. The 20-second evaluation in Table 1 and Figure 3 provides one data point on generalization. There are no experiments at other durations, no analysis of how VBench scores or drift metrics vary with generation length, and no investigation of information propagation beyond the attention window.

Mitigation status. The paper does not acknowledge the limited scope of the generalization evaluation or discuss what would be needed to characterize performance at longer horizons. The sliding-window design is described as a practical implementation detail (Section 4.1), not as a potential limitation for very long generation. The paper's strong claims about indefinite-length generation are backed by evidence that the generation remains stable at 4× the training duration—an encouraging result, but far from what is needed to support "indefinite extension."

---

6.6 Inference-Only Speedup Depends on Specific Step Count and May Not Scale

The assumption or constraint. The $\sim 1.8\times$ wall-clock speedup (30 fps vs. 17 fps, Table 1) is demonstrated for a specific configuration: $S = 4$ denoising steps, with the pipelined parallelism and KV-cache fusion optimizations described in Section 3.2. The speedup arises from processing the $N \times S$ grid along anti-diagonals with $N + 2$ fused forward passes per step rather than $2N$ unfused passes, plus parallelism across denoising steps.

The paper does not report how the speedup scales with the number of denoising steps $S$, the number of blocks $N$, or the hardware configuration. The 4-step schedule is unusually aggressive for diffusion models—most production video diffusion systems use 20–50 denoising steps for quality reasons, and the paper's own teacher model (used for DMD critic training) runs 50 ODE steps for trajectory generation. The distillation to 4 steps is a specific design choice that trades quality for speed; if an application required more denoising steps (e.g., 8 or 16) for higher quality, the pipelining benefit might change substantially.

The consequence. The practical applicability of the speedup depends on whether the use case can tolerate a 4-step schedule. If not, the headline throughput number may not be achievable while maintaining the reported quality:

Scaling with denoising steps. With $S$ steps, the $N \times S$ grid has $S$ pipeline stages. When $S$ is small (4 steps), the pipeline is shallow—there are only 4 stages to parallelize across, and the pipeline startup and drain phases (where stages are idle waiting for data) represent a significant fraction of total execution time. As $S$ increases, the pipeline deepens, and the theoretical utilization approaches 100%. However, communication overhead also scales with $S$ (more inter-stage messages), and the KV-cache fusion optimization (reducing $2N$ passes to $N + 2$ per step) is independent of $S$—its benefit is per-step, not cumulative across steps. The paper provides no scaling curve to characterize this trade-off.

Fusion benefit degrades with larger context windows. The KV-cache fusion optimization concatenates context and target blocks along the frame dimension with different per-frame timesteps (Section 3.2). This requires that the concatenated sequence fit within the model's maximum sequence length (determined by the DiT architecture's positional encodings and attention memory). If the context window were expanded beyond 5 seconds, or if blocks contained more frames, the concatenated sequence might exceed this limit, requiring a fallback to the unfused $2N$ passes per step. The paper does not discuss this constraint or report the maximum concatenated sequence length in their implementation.

Hardware-dependent speedup. The pipelined parallelism described in Section 3.2 assumes multiple processes (or GPUs) dedicated to different denoising steps, with asynchronous point-to-point communication. The speedup is measured relative to other distilled AR models (CausVid, Self-Forcing, Causal Forcing) that "share the same Wan2.1-1.3B backbone and 4-step denoising schedule." It is unclear whether the baseline models implement optimal batching, KV-caching, or other optimizations available to them—the speedup might partially reflect engineering differences rather than a fundamental throughput advantage of the hierarchical order. The paper does not report the hardware used, the number of GPUs, the communication protocol, or whether the baseline models were optimized by the same team.

What evidence exists in the paper. Table 1 reports 30 fps vs. 17 fps for HiAR vs. other distilled models. Section 3.2 describes the pipelining and fusion optimizations. No scaling experiments (varying $S$, $N$, or hardware) are reported.

Mitigation status. Not addressed. The paper presents the speedup as a fixed property of the architecture (Abstract: "This hierarchy naturally admits pipelined parallel inference, yielding a ∼1.8× wall-clock speedup in our 4-step setting") without discussing how it generalizes to other configurations. The qualification "in our 4-step setting" is present but not elaborated—the reader is not told what to expect at 8 steps, 16 steps, or with different hardware.

7. Implications and Future Directions

How This Work Changes the Landscape

This paper causes a conceptual reframing of autoregressive video generation, not by proposing a new model architecture or a new loss function, but by identifying that the generation order—the sequence in which blocks and denoising steps are interleaved—is a first-class design dimension with profound consequences for error propagation, motion diversity, and inference efficiency. Prior to this work, the block-first generation order was essentially invisible as a design choice. It was inherited from the language modeling paradigm where tokens are generated sequentially and completely before moving to the next, and it seemed intrinsic to what "autoregressive" means. HiAR demonstrates that this conflation of causal dependency with execution order is a category mistake: causal structure specifies what depends on what, not the order in which those dependencies are resolved.

The magnitude of this shift is substantial but bounded. It is not a paradigm shift on the scale of the diffusion-to-AR transition—the underlying diffusion framework, the DiT backbone, and the DMD distillation objective are all inherited from prior work. Rather, it is a reorganization of the computation graph within the existing paradigm that exposes new trade-offs and new optimization opportunities. The paper shows that this reorganization, when paired with appropriate training and regularization, improves quality (VBench Total 0.821 vs. 0.805 for Self-Forcing), reduces drift (0.257 vs. 0.355, a 27.6% reduction), preserves motion diversity (Dynamic 0.686 vs. 0.542), and improves throughput (30 fps vs. 17 fps)—all from changing the order of operations while holding the model architecture, training data, and inference compute budget constant. Improvements that span quality, stability, diversity, and efficiency simultaneously are rare, and they signal that the generation order was an under-explored lever.

Resolution of a prior contradiction. The paper reconciles an apparent tension in the video generation literature. Bidirectional models produce high-quality, temporally coherent videos but are limited to fixed durations. Autoregressive models can scale to arbitrary lengths but suffer progressive quality degradation. The field had implicitly accepted this as an inherent trade-off: scalability comes at the cost of stability. HiAR shows that this trade-off is not inherent to autoregressive generation per se, but is an artifact of the clean-context convention that all prior AR methods adopted. By providing context at matched noise levels rather than fully denoised, HiAR achieves AR scalability while approaching bidirectional-level temporal stability—the Drift score of 0.257 represents a regime where degradation is dramatically reduced but not eliminated, and the Quality score of 0.846 actually exceeds the bidirectional teacher's 0.813. This reframes the problem from "AR generation is inherently unstable" to "AR generation with clean context is unstable," opening the possibility that other context noise schedules or generation orders might further narrow or close the remaining gap.

Directions that become more attractive. The paper's central insight—that context noise levels control a bias–information trade-off, and that the optimal level is the noisiest that still satisfies temporal causality—has implications beyond video. Any sequential generation process where earlier outputs condition later ones might benefit from matched-noise conditioning: autoregressive image generation (generating image patches sequentially), autoregressive audio generation, or even autoregressive language modeling if recast in a diffusion framework. The hierarchical generation order (step-first rather than block-first) might also generalize: any multi-step sequential generation process can be viewed as an $N \times S$ grid, and traversal orders other than row-major may expose parallelism or modulate error propagation. The paper's demonstration that generation order cannot be changed at inference time without retraining (Table 3, "w/o re-training": Quality drops to 0.767) means this is a training-time architectural decision, not a post-hoc optimization, which elevates its importance for model designers.

Directions that become less attractive. The paper's results implicitly de-emphasize two lines of work. First, the finding that matched-noise context outperforms clean context at all tested budgets suggests that methods focused solely on improving the accuracy of previous-block predictions (e.g., better teacher forcing, more accurate rollouts) without addressing how those predictions are presented as context are optimizing the wrong objective. Even a perfect prediction of the previous block, if presented at $t_c = 0$, would propagate whatever residual error exists at full strength. Second, the paper's demonstration that lookahead search and more aggressive optimization can paradoxically hurt performance due to verifier over-optimization (the reference example's finding, which parallels HiAR's finding that more forward-KL steps degrade quality in Table 3) suggests that complex search or constraint mechanisms layered on top of standard AR generation may be less promising than fundamentally rethinking the generation order and context presentation. Simpler architectures with smarter information flow may outperform more complex architectures with conventional information flow.

A new diagnostic capability. The paper introduces a quantitative drift metric that captures temporal degradation through the slope of per-segment statistics over a 20-second video. While the metric itself requires validation (see Section 6.4), the concept of measuring drift as a rate rather than an aggregate quality score is an important methodological contribution. It enables the field to move beyond qualitative observations of "oversaturation over time" or "progressive quality degradation" to quantitative comparisons between methods, and it surfaces temporal stability as a first-class evaluation dimension alongside average quality. This is analogous to how the field moved from reporting aggregate BLEU scores to measuring position-specific degradation in machine translation—the shift from "how good is the output on average" to "how does quality change over the course of the output" is conceptually significant even before the specific metric is perfected.

Follow-Up Research This Work Enables

Characterizing the empirical scaling of drift with generation duration. The paper demonstrates drift reduction at 20 seconds (4× the 5-second training duration), but provides no evidence about how drift evolves at longer horizons—40 seconds, 60 seconds, 5 minutes, or longer. A strong follow-up would measure the Drift Score and per-segment VBench scores at multiple generation durations (10 s, 20 s, 40 s, 80 s, 160 s) for HiAR, Self-Forcing, and Causal Forcing, fitting a scaling law of the form $\text{Drift}(T) = \alpha \cdot T^\beta$ or $\text{Drift}(T) = \alpha \log(T) + c$ to determine whether drift accumulates linearly, sub-linearly, or super-linearly with duration. This would answer the central open question from Section 6.3: does the matched-noise attenuation merely reduce the drift rate by a constant factor (in which case indefinite-length generation is still fundamentally limited), or does it change the functional form of error accumulation (in which case very long generation might be practical)? The experiment would require generating videos up to several minutes—well beyond the paper's current evaluation—and would stress-test the sliding-window KV cache design. If drift accelerates at long horizons, it would motivate research into explicit long-horizon memory mechanisms beyond the 5-second attention window. If drift remains linear or sub-linear, it would validate the paper's implicit claim that hierarchical denoising enables indefinite-length generation.

Direct causal validation of the matched-noise error attenuation mechanism. The paper's theoretical analysis in Section 3.1 predicts that matched-noise context reduces inter-block error propagation by the attenuation coefficient $(1 - \sigma_{t_c})$, but this causal mechanism is never directly tested—the empirical evidence is correlational (lower drift at $t_c = t_{j+1}$ vs. $t_c = 0$). A direct experiment would inject synthetic prediction errors of known magnitude into the context for block $B_{n-1}$ and measure how these errors propagate to block $B_n$ at different context noise levels. Specifically: take a ground-truth video, generate block $B_{n-1}$ normally, add a controlled perturbation $\delta$ to its clean prediction (e.g., Gaussian noise scaled to a specific SNR), construct context at various $t_c$ levels using Equation 6, and measure the resulting error in block $B_n$ compared to ground truth. The key prediction is that the propagated error should scale as $(1 - \sigma_{t_c}) \cdot \|\delta\|$—the attenuation coefficient from Equation 7. This experiment would also reveal whether the model exhibits nonlinear error amplification (errors larger than predicted by the linear decomposition) or error suppression (the model partially corrects for known error patterns), which would refine or challenge the simple linear bias–information model. A negative result—finding that error propagation does not follow the predicted attenuation pattern—would suggest that the drift reduction from matched-noise context operates through a mechanism other than the one proposed, motivating a revised theoretical framework.

Testing hierarchical denoising on a different backbone architecture and video domain. All experiments in the paper use Wan2.1-1.3B. A critical stress-test would replicate HiAR on a substantially different video diffusion architecture. Candidates include: (1) a UNet-based video diffusion model (e.g., Stable Video Diffusion, Blattmann et al., 2023) to test whether the matched-noise principle depends on the transformer attention mechanism; (2) a model with a different noise schedule (different shift parameter $s$ in Equation 1) to test whether the optimal $t_c = t_{j+1}$ result is sensitive to the schedule curvature—the paper's derivation assumes only that SNR is monotonic in $t$, but the specific attenuation coefficient $(1 - \sigma_{t_c})$ depends on the schedule; (3) a non-distilled AR model trained from scratch with hierarchical denoising (without the DMD + forward-KL distillation pipeline) to determine whether the benefits are tied to the distillation paradigm or are inherent to the hierarchical generation order. The experiment would also test on a non-VBench dataset—for instance, a driving-scene dataset where temporal stability is critical for downstream tasks like trajectory prediction, or an egocentric video dataset where motion diversity is naturally high—to distinguish VBench-specific effects from general properties. If hierarchical denoising transfers cleanly, it establishes the matched-noise principle as architecture-agnostic. If it fails on certain architectures or domains, it reveals boundary conditions that would refine the theory.

Combining hierarchical denoising with diffusion forcing training. The paper's "w/o re-training" ablation (Table 3) tests the extreme case: a model trained with block-first clean-context (Self-Forcing) applied to hierarchical denoising at inference. An intermediate case would be informative: train with diffusion forcing (Chen et al., 2024), which exposes the model to heterogeneous context noise levels during block-first training, and then test whether this partially bridges the train–test gap when hierarchical denoising is applied at inference. The hypothesis is that diffusion forcing's noisy-context training provides some robustness, potentially reducing the quality penalty from 0.767 to something closer to 0.82–0.83, making retraining less critical or enabling a shorter fine-tuning phase rather than full retraining. The experiment would use the same Wan2.1-1.3B backbone, replace the Self-Forcing pretraining with diffusion forcing pretraining (varying per-token noise levels during block-first training), and then evaluate with hierarchical denoising at inference both with and without additional fine-tuning under the hierarchical schedule. A positive result (diffusion forcing partially closes the gap) would suggest a practical pathway for adapting existing models to hierarchical denoising without full retraining. A negative result (diffusion forcing provides no benefit over Self-Forcing for hierarchical inference) would indicate that the train–test mismatch is not primarily about noise-level robustness but about the fundamental difference between block-first and step-first information flow, requiring the model to learn qualitatively different temporal reasoning patterns.

Investigating whether the forward-KL regularizer prevents collapse indefinitely or merely delays it. Table 3 shows that without the forward-KL regularizer, dynamics collapse to 0.445 after 20,000 training steps under the hierarchical schedule. With the regularizer ($K = 1$, $\lambda = 0.1$), dynamics are maintained at 0.686. A critical open question is whether this protection is permanent or merely extends the collapse horizon. The follow-up experiment would extend training to 40,000, 80,000, or 160,000 steps (if computationally feasible) and track Dynamic scores throughout. The key measurement is whether the Dynamic score asymptotes at some stable value (indicating the regularizer creates a new equilibrium) or slowly declines (indicating the regularizer only slows the collapse, and the mode-seeking reverse-KL objective eventually dominates). This has direct practical implications: if collapse is merely delayed, practitioners must carefully tune the number of training steps to stop before degradation begins, and research should focus on stronger or fundamentally different regularization. If collapse is prevented indefinitely, the forward-KL regularizer at $K = 1$, $\lambda = 0.1$ is a robust recipe. Additionally, measuring the Dynamic score of the forward-KL-only model (trained without DMD loss) at long horizons would characterize the baseline motion diversity achievable through trajectory matching alone, establishing an upper bound on what the regularizer can preserve.

Extending the hierarchical denoising principle to variable-length and interactive generation. The paper evaluates HiAR on fixed-length 20-second generation from text prompts. A natural extension is to settings where video length is not predetermined and where the generation process must respond to external inputs—precisely the "interactive agents and world models" applications the paper cites in Section 1. In an interactive setting, the model receives new conditioning signals (e.g., user actions, environmental changes) at arbitrary points during generation and must continue coherently. The hierarchical denoising framework could be adapted so that when new conditioning arrives at block $k$, the model halts, re-initializes or adjusts the remaining blocks, and continues the step-first sweep. The research question is whether matched-noise context provides sufficient flexibility for such interventions—can a block at noise level $t_j$ be meaningfully redirected by new information at that noise level, or does effective intervention require cleaner context? The experiment would implement a simple interactive loop: generate a base video for 10 seconds, inject a new text prompt at 5 seconds, and measure whether the video transitions coherently to the new prompt while maintaining visual quality. Comparisons against block-first AR models (which naturally handle prompt changes at block boundaries) would reveal whether hierarchical denoising's interleaved noise levels help or hurt responsiveness.

Practical Applications and Downstream Use Cases

Real-time streaming video generation for interactive applications. The combination of 30 fps throughput and 0.30 s per-chunk latency (Table 1) makes HiAR viable for applications requiring continuous, low-latency video output—interactive game environments, real-time video avatars, or streaming world models where an agent's actions must be reflected in the visual output within a fraction of a second. The 0.30 s per-chunk latency means a new 3-frame chunk arrives every 300 ms, which is within the acceptable range for interactive responsiveness (typical game rendering targets 30–60 fps, corresponding to 16–33 ms per frame; HiAR's per-frame latency is approximately 100 ms for 3 frames). The 5-second sliding-window KV cache keeps memory bounded regardless of session duration, preventing the cost-per-frame from growing over time—a critical property for deployment scenarios where sessions might last minutes or hours. A video avatar system could use HiAR to generate a talking-head video stream where the model conditions on both previous frames (for temporal coherence) and a streaming audio or text input (for lip-sync and expression), with the hierarchical denoising providing stable visual quality over extended conversations. The 4× extrapolation from 5-second training clips to 20-second generations demonstrated in the paper would need to be extended to much longer horizons for practical deployment, but the architecture's bounded-memory design is already suited to indefinite-length operation, pending validation of stability at longer durations.

Cost-efficient distillation of large video models for edge deployment. HiAR demonstrates that a 1.3B-parameter student model, distilled from a 14B-parameter teacher and trained with hierarchical denoising, can outperform the teacher's bidirectional generation on VBench Quality (0.846 vs. 0.813) while running at 30 fps—nearly 40× faster than the teacher's 0.78 fps. This has direct implications for deploying video generation on consumer hardware (laptops, phones, VR headsets) where the 14B model would be infeasible. The distillation procedure—16,000 ODE pairs for DMD training, 20,000 trajectories for forward-KL regularization, 20,000 training steps at batch size 64—is computationally intensive but represents a one-time cost amortized across all downstream inferences. An organization with a large proprietary video model could use HiAR's recipe to produce a lightweight, real-time version for client-side deployment, with the hierarchical generation order providing both the quality and speed benefits documented in the paper. The specific numbers from Table 1 provide a benchmark: a 1.3B model at 30 fps achieves VBench Total 0.821, compared to the 14B teacher's 0.802 (bidirectional, 0.78 fps). The quality improvement over the teacher (0.821 vs. 0.802) is an unusual property—distillation typically trades quality for speed, not improves it—and suggests that the hierarchical denoising training procedure has a regularizing or denoising effect beyond what the teacher achieves natively.

Long-form video content creation with consistent visual quality. For applications generating videos of 20 seconds to several minutes—marketing content, educational videos, synthetic data for training perception systems—the primary failure mode of current AR methods is progressive degradation: colors shift, motion becomes repetitive, and details blur over time. HiAR's 27.6% drift reduction relative to Self-Forcing (0.257 vs. 0.355) means that a 20-second generated video retains visual quality substantially closer to its starting point than competing AR methods. In a content creation pipeline where a human operator selects the best of several generated videos or applies post-processing, HiAR's stability means fewer generations need to be discarded due to late-sequence degradation, reducing the total compute cost per usable output. The forward-KL regularizer's motion diversity preservation (Dynamic 0.686 vs. Self-Forcing's 0.542) further means that generated videos maintain natural-looking motion throughout their duration rather than becoming progressively static—a failure mode that is particularly noticeable to viewers and difficult to fix in post-processing. A content creation tool could expose a "duration" slider to users, with HiAR generating stable video at any requested length, where competing AR methods would show visible quality falloff beyond 10–15 seconds and bidirectional methods would be capped at their fixed training window (typically 5 seconds).

When to Prefer This Method

The paper positions HiAR explicitly against other distilled autoregressive methods (Self-Forcing, Causal Forcing, CausVid) and implicitly against bidirectional diffusion models, though the generation paradigms serve different use cases. The decision rules below are grounded in the paper's empirical comparisons from Table 1 and the ablations in Tables 2–3.

• Prefer HiAR over other distilled AR methods when motion diversity and temporal stability are both critical. Self-Forcing achieves competitive Quality (0.829 vs. HiAR's 0.846) but substantially lower Dynamic (0.542 vs. 0.686) and higher Drift (0.355 vs. 0.257). Causal Forcing closes the Dynamic gap somewhat (0.672) but at higher Drift (0.615). HiAR is the only distilled AR method that simultaneously approaches the teacher's motion diversity (0.686 vs. 0.690 for Wan2.1-1.3B) and achieves the lowest drift. If the application involves human subjects, natural scenery with expected motion (wind, water, walking), or any content where unnatural stillness would be perceptually obvious, the forward-KL regularizer's motion preservation is likely worth the additional training complexity.

• Prefer HiAR over other distilled AR methods when inference latency and throughput matter. All distilled AR models share the same backbone and denoising steps, but HiAR's pipelined parallelism yields a ~1.8× throughput advantage (30 fps vs. 17 fps) and corresponding latency reduction (0.30 s vs. 0.69 s per chunk). If the deployment scenario involves real-time or interactive applications, or if the total cost of serving many requests is dominated by inference time, HiAR's structural efficiency advantage is compelling, and it comes without sacrificing quality—in fact, HiAR simultaneously achieves the best VBench scores.

• Prefer bidirectional models (Wan2.1, Sora, etc.) when generation is limited to short clips (≤5 seconds) and no autoregressive extension is needed. Bidirectional models remain the quality baseline for fixed-duration generation within their training window. HiAR's Quality score exceeds Wan2.1-1.3B (0.846 vs. 0.813), but this comparison is confounded by the evaluation length mismatch (bidirectional models evaluated at 5 s, HiAR at 20 s). If the application only ever needs 5-second clips and never requires streaming, interactive, or indefinite-length output, the simpler bidirectional pipeline—no self-rollout training, no forward-KL regularization, no KV-cache management—may be preferable for engineering simplicity despite slightly lower reported quality.

• Prefer non-distilled AR models (MAGI-1, SkyReels-V2) only if inference speed is irrelevant and model scale is the primary quality driver. The non-distilled AR models in Table 1 operate at substantially lower throughput (0.19–0.49 fps) and higher latency (112–282 s), making them impractical for real-time or interactive use. Their VBench scores are also lower than HiAR's (Total 0.757–0.788 vs. 0.821), but this comparison is confounded by their higher denoising step counts—they trade speed for per-step quality in a way that the 4-step distilled models do not. If a use case can tolerate minutes of generation time per video and prioritizes per-frame fidelity over motion diversity and temporal stability, the non-distilled models may be worth evaluating, though the paper's evidence does not directly support this trade-off since HiAR outperforms them on all reported metrics.

• Do not use hierarchical denoising at inference only (without retraining). Table 3 ("w/o re-training") shows Quality drops to 0.767 and Semantic to 0.559—the worst visual quality of any configuration tested. The train–test mismatch from changing the generation order without corresponding training is severe and not a viable deployment strategy. If retraining is infeasible, Self-Forcing or Causal Forcing are preferable to inference-only hierarchical denoising.