acceptodds
Under review as a conference paper at ICLR 2027

Distilling Uniform Discrete Diffusion Models into a One-Step Video Generator

Abstract

Discrete diffusion has emerged as a competitive paradigm for visual generation, yet high-quality generation still relies on repeated refinement steps, limiting its use in low-latency interactive video generation. Compressing this process into a single forward pass is particularly challenging for video, where tens of thousands of categorical variables must jointly establish coherent spatial structure and temporal dynamics without subsequent correction. In this work, we present DDMD-V, to our knowledge the first one-step distillation framework for uniform discrete diffusion video generation. DDMD-V extends on-policy distribution matching to the uniform discrete diffusion setting, using student-induced intermediate states along the teacher’s metric-dependent corruption path. To complement token-level distribution matching, we further introduce spatiotemporal adversarial regularization in the FSQ embedding space, providing a learned signal over multi-token spatial and temporal configurations. Distilled from URSA-1.7B, our student generates a 49-frame \(320\times512\) video comprising discrete tokens in a single network forward pass. It achieves 81.78 VBench score compared with the 81.75 score of the 50-step teacher, while reducing the number of function evaluations (NFEs) from 100 to one when counting the teacher's conditional and unconditional guidance branches. After distillation, the student achieves closely matched visual quality compared to the teacher. In contrast, directly reducing the teacher to one sampling step yields only 58.56 VBench. With BF16 inference, DDMD-V reaches a generation throughput of more than 50 video frames per second on a single H200 GPU. These results demonstrate that a strong discrete video diffusion model can be distilled into a one-step generator with comparable performance, providing a promising path toward low-latency generative video models. %for interactive world-model systems. We will release the code and model.

Primary area:
generative models
Keywords:
Discrete Diffusion, One-step Distillation, Video Generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_rjW9OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents DDMD-V, described as the first one-step distillation framework for uniform discrete diffusion video generation. The teacher, URSA-1.7B, generates 49-frame 320×512 videos (N = 33,280 FSQ tokens, vocabulary K = 64,000) via a metric-dependent probability path requiring ~50 sampling steps (100 NFEs counting the conditional/unconditional CFG branches). DDMD-V distills this into a student that produces the entire token grid in a single forward pass from a uniform categorical source.

Critical strengths
  1. A striking, well-isolated result on an important problem. Matching a 100-NFE teacher (81.78 vs. 81.75 VBench) with a single forward pass and a measured 77.6× wall-clock speedup is a substantial efficiency result for low-latency video generation. It is consequential because the paper rules out the trivial explanation: truncating the teacher to one step collapses to 58.56, so the gain is genuinely attributable to distillation rather than to the task being easy. The result is further corroborated by a three-judge (human + two model) preference study, which moves the headline from "one metric went up" to "quality is broadly comparable." This is the primary reason I lean toward acceptance.

  2. Design choices are mechanistically motivated and empirically validated, not just asserted. The two signals target a specific, well-identified failure mode — a one-step generator cannot correct multi-token spatiotemporal structure after the fact — and the ablations substantiate each: removing the adversarial critic costs 3.51 VBench (with per-dimension degradation concentrated in subject/background consistency and flickering), and restricting corruption-time coverage to the entropy band collapses the model to 74.63. This tight coupling between claimed mechanism and measured effect materially increases my confidence that the method works for the stated reasons.

Critical weaknesses
  1. Generality rests on a single teacher at a single scale. Every result is obtained by distilling URSA-1.7B. The paper claims to demonstrate the feasibility of one-step distillation for uniform discrete video diffusion as a paradigm, but with n = 1 teacher and one model size, it is impossible to tell whether the recipe (corruption-time coverage, η = 0.5 bidirectional KL, adversarial weight, auxiliary update ratio) transfers or was tuned to this specific teacher. For a framework-level contribution, this is the single biggest limitation on how far the conclusions generalize.

  2. No comparison against competing distillation methods adapted to this setting. The only comparators are the teacher at different sampling budgets. The related work cites Di[M]O, Soft-Di[M]O, and discrete MMD distillation, but none is adapted to the video / uniform-discrete setting as a baseline. Consequently the paper establishes that DDMD-V vastly outperforms truncating the teacher, but not that it is competitive with — let alone preferable to — alternative distillation paradigms. Without this, the specific design's advantage over simpler or existing distillation approaches is unestablished, and the "first + best" framing rests only on the former.

Questions for authors

(→ W1, generality) Can you demonstrate transfer beyond URSA-1.7B? Concretely, apply the same hyperparameters to at least one additional uniform-discrete teacher and/or a second model scale, and report whether the recipe holds without per-teacher retuning. If retuning is required, please report the sensitivity of VBench to the key knobs (corruption-time coverage, η, adversarial weight, auxiliary-to-student update ratio) so readers can judge how teacher-specific the method is.

(→ W2, missing distillation baselines) Please add at least one competing distillation baseline adapted to this setting — e.g., a Di[M]O/Soft-Di[M]O-style objective or discrete-MMD distillation applied to URSA video — trained under identical data, cache, and compute budgets. The key deliverable is evidence that DDMD-V is competitive with or superior to existing distillation paradigms, not only to teacher-sampler truncation.

(→ "comparable" claim / VBench saturation) Given that 81.78 vs. 81.75 is within VBench's effective noise and that the preference study shows only approach to parity with the 50-step teacher, can you report the VBench variance across seeds/evaluation runs and the confidence intervals on the 50-step preference comparison? A quality metric less prone to saturation (and any observed failure modes relative to the teacher) would make the "comparable performance" claim more defensible.

(→ clarification, adversarial dependence) The adversarial branch relies on a cache of 3,200 teacher-generated videos. How sensitive is final quality to cache size and to the ratio of real-teacher to student fields, and is there any risk the student is partly matching the cached set rather than the teacher distribution? Reporting a cache-size sweep would clarify whether the adversarial signal is a lightweight regularizer or a load-bearing dependency on teacher samples.

3 comments

We thank the reviewer for positive assessment. We address each concern below.

W1 / Q1. Generality rests on a single teacher at a single scale. Can you demonstrate transfer beyond URSA-1.7B?

A1. This is a fair limitation and we now state it explicitly. To our knowledge, URSA is the only publicly available family of uniform discrete video diffusion models, so a teacher from a different family is not available. We distilled the second released video checkpoint, URSA-0.6B, using the same training recipe:

TeacherTeacher, 50 stepsTeacher, 1 stepDDMD-V, 1 step
URSA-1.7B (paper)81.7558.5681.78
URSA-0.6B (new)80.9257.3581.14

Moreover, we'd like to emphasize that the DDMD-V objective relies only on the teacher's forward corruption kernel and its clean-token posterior, both of which are defined for any uniform discrete diffusion model. Nothing is specific to URSA's architecture or tokenizer. We hope this can address the concern.

W2 / Q2. No comparison against competing distillation methods adapted to this setting. Please add at least one competing distillation baseline trained under identical data, cache, and compute budgets.

A2. We implemented Di[M]O, Soft-Di[M]O and trained them under the same setting. The results are shown below:

MethodVBench Total
Teacher (1 step)58.56
Di[M]O-style74.59
Soft-Di[M]O-style78.69
DDMD-V81.78
Teacher (50 steps)81.75

Q3. Given that 81.78 vs. 81.75 is within VBench's effective noise, can you report the VBench variance across seeds and the confidence intervals on the 50-step preference comparison?

A3. VBench. We performed paired bootstrap resampling over the 946 prompts (B = 2,000):

ModelTotal95% CI
Teacher, 50 steps81.75[81.07, 82.34]
Ours, 1 step81.78[81.27, 82.25]
Paired difference (ours − teacher)+0.03[−0.44, +0.55]

Re-seeding the student's evaluation gives 81.78 / 81.83 / 81.76 (mean ± SD: 81.79 ± 0.04). We also report an updated student model. It scores 82.25 / 82.36 / 82.57 (82.39 ± 0.16) over three seeds, with a paired difference to the teacher of +0.51 (95% CI [−0.17, +1.21]).

Preference comparison. We further compute 95% confidence intervals for the preference differences against the 50-step teacher using bootstrap resampling over evaluation prompts (B = 2,000):

EvaluatorTeacher / Tie / OursΔ Preference95% CI
Human Raters25 / 48 / 27%+2 pp[−2.4, +6.4]
GPT-6 Astra46 / 8 / 46%0 pp[−6.0, +5.9]
Gemini 3.8 Flash35 / 23 / 42%+7 pp[+1.6, +12.3]

Additional quality metric. We present additional quality evaluation on VideoReward in App. F.1 Tab. 5. Our student is also closely matching the full 50-step teacher.

Observed limitations and failure modes. We retrained DDMD-V and evaluated 10 checkpoints saved during the final 500 training updates.

MetricTeacher, 50 stepsOurs, submittedCheckpoints above teacher
Semantic76.8876.100/10
Subject95.7996.239/10
Action93.6097.0010/10
Color79.8083.719/10
Aesthetic62.1060.530/10
Imaging62.2359.100/10

Subject, Action, and Color consistently outperform the 50-step teacher, whereas Semantic, Aesthetic Quality, and Imaging Quality remain lower. Notably, Semantic and Imaging Quality generally fall between those of the 16-step and 32-step teachers.

The primary limitation is lower per-frame visual fidelity. Increasing the teacher's sampling steps from 16 to 50 improves Imaging Quality by 3.43 points and Aesthetic Quality by 1.01 points, suggesting that later sampling steps mainly refine fine-grained visual details. One-step generation sacrifices some of these refinements while maintaining strong spatiotemporal consistency.

Taken together, these evaluations support our central claim of comparable overall quality with substantially fewer sampling steps, while acknowledging the remaining gap in fine-grained visual fidelity.

Q4. The adversarial branch relies on a cache of 3,200 teacher-generated videos. How sensitive is final quality to cache size and to the ratio of real-teacher to student fields, and is there any risk the student is partly matching the cached set rather than the teacher distribution?

A4. We address the memorization concern first, then the sensitivity.

  1. No evaluation prompt overlaps with the cache. The cache prompts are long descriptive captions from Koala-36M, whereas evaluation uses VBench prompts. We checked for overlap and found none of the 943 VBench prompts appears among the cache prompts.

  2. No evidence of direct memorization. We further examined whether the student reproduces cached teacher videos. Using the submitted model, whose critic was trained with 3,200 cached videos, we compared student and teacher outputs under the same (c,ξ)(c, \xi). Token agreement was 0.029% (± 0.015%) over 300 cached pairs, compared with 0.067% (± 0.046%) over 300 unseen teacher pairs. Agreement on cached pairs was not higher than on unseen pairs, and both remained below 0.1%, providing no evidence of direct cache memorization.

  3. Cache-size and sampling-ratio sensitivity. We ablate with different cache sizes and with different ratios of real-teacher to student fields using 50 / 100 steps training:

Cache size1,6003,2009,600
VBench Total80.18 / 80.8580.12 / 80.8080.11 / 80.95
real-teacher : student fields1 : 41 : 21 : 1
VBench Total80.31 / 81.0780.11 / 80.9580.10 / 80.60

The results indicate both the cache size and the sampling-ratio are not sensitive to the final quality.

  1. Regularizer or load-bearing dependency? The two parts should be separated. The spatiotemporal critic itself is load-bearing: removing it lowers the result clearly (App. H.1). The cached teacher samples are not: the critic needs them only as examples of the teacher’s output distribution, and a cache two times smaller gives the same result.

The authors have answered all of my questions, but the overall quality of this paper does not justify a score of 4, so I have decided to maintain my original score.

Reviewer_ELcfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed the first one-step distillation method for uniform discrete video diffusion model. It points out the gap between discrete and continuous diffusion model lies in the categorical video tokens, and proposed mechanisms targeted at this issue. The proposed distillation is built upon DMD loss and adversarial loss: The distribution-matching loss ensures that in the gradient can pass through the non-continuous categorical tokens to match the teacher distribution and the recovered auxiliary distribution, which is trained to match the student distribution; the adversatial loss ensures that teacher generated examples and student generated ones are indistinguishable, further improving spatiotemporal consistency. The experiments show that the method recovers the performance of the teacher and matches other non-distilled models.

Critical strengths
  • The method proposed solves a very important missing piece in video diffusion distillation literature, and helps bring down the cost for discrete video diffusion, which is an increasingly important paradigm.
  • Observations specific to discrete video diffusions, like how the low-noise region behaves, are different from continuous video diffusion and are thus provide valuable insight.
  • The method proposed is easy to understand and straightforward while providing good performance.
Critical weaknesses

Despite the importance of the question, some of the designs here are straightforward extensions of the continuous case (e.g. extending continuous loss to categorical, using adversarial loss for spatiotemporal consistency) or the setting of images (e.g. the construction of differentiable token-embedding fields), so part of the method is not that novel.

Questions for authors
  • The observation regarding low-noise region in the last ablation in 4.4, which is also reiterated in intro line 107, is interesting to me, which I find could be an important distinction between the discrete and continuous case. However, there's no discussion or explanation of why this is the case. While it's standard practice that people reweight the SNR so that high noise region is sampled more, clipping out a large portion of the low noise region completely doesn't seem to be very common to my knowledge, which might indicate a fundamental gap that inhibit the direct migration of the continuous case to the discrete case. I think the novelty of the paper will increase if more explanation and investigation on the underlying mechanism could be discussed.

  • Can you elaborate on the necessity of the augmentation and sharing it across teacher and student latents in line 301-302? The motivation of randomly replacing with random token and the shortcut it aims to fix is not clear. Even if the given augmentation does makes real and generated videos indistinguishable in undesired aspects we want to avoid (which I'm not sure what they are), since the mask is binary, each token in the real latent always belong to the codebook while the student's latent can be interpolation of the embeddings. This implies that even with such augmentation, shortcut still exists. Could the author explain why this shortcut isn't an issue?

5 comments

We thank the reviewer for the two pretty insightful questions. We address each concern below.

W1. The method novelty should be positioned precisely. A controlled adaptation baseline would better isolate these video-specific improvements.

A1. We thank the reviewer for recognizing our paper addresses an important missing piece. We would like to explain why DDMD-V is not a straightforward extension of existing methods. DDMD-V is the first method that distills a uniform discrete diffusion video model into a one-step generator, and achieving comparable performance to its teacher is non-trivial. Without spatiotemporal critic, the student can not achieve performance comparable to that of the teacher. Without the corruption-time restriction, distillation fails when the distribution-matching and adversarial losses are applied jointly from the first update.

We implemented Di[M]O, Soft-Di[M]O, DDMD-V without the corruption-time restriction on URSA and trained them under the same setting. The results are shown below:

MethodVBench Total
Teacher (1 step)58.56
Di[M]O-style74.59
Soft-Di[M]O-style78.69
DDMD-V w/ full tt range52.59
DDMD-V w/ t∈[0.05,0.65]t \in [0.05, 0.65]81.78
Teacher (50 steps)81.75

The Di[M]O-style and the Soft-Di[M]O-style students reach 74.59 and 78.69, respectively, against 81.78 for DDMD-V. Moreover, without the proposed corruption-time restriction, DDMD-V collapses to prompt-independent outputs within the first 100 training steps.

Our contributions: First, uniform discrete diffusion has no mask token, so the initialization and mask-ratio schedules of masked image distillation do not apply. Instead, we re-corrupt the student's own samples along the teacher's metric-dependent path and find that appropriately covering the heavily corrupted part is essential. Second, a single pass must fix the joint structure of 33,280 tokens across 49 frames and a spatiotemporal critic achieves this.

"Simplicity is the ultimate sophistication." We would also like to emphasize that we explored and experimentally evaluated several alternative designs that ultimately proved unnecessary. Every component we kept is one whose removal causes a measured loss. Components that did not help were removed, for example an explicit metric-aware matching kernel (App. H.4) and straight-through critic inputs (App. H.5).

Q1. The observation regarding the low-noise region is interesting, but there is no explanation of why this is the case. Clipping out a large portion of the low-noise region completely does not seem very common, which might indicate a fundamental gap.

A2. Thank you for highlighting this important observation. We revisited this ablation and conducted a more detailed analysis.

Our key finding is: The low-noise region is not inherently harmful; its effect depends on the training schedule. We observe two distinct behaviors:

  • Staged training (as in our submission): When the critic is introduced after distribution matching warm-up, extending the corruption interval to the full range during the second stage has little effect on performance.

  • Joint training (critic enabled from the start): Training with the full corruption range collapses to prompt-independent outputs within the first 100 steps, whereas restricting the interval to t∈[0.05,0.65]t \in [0.05, 0.65] enables stable optimization.

Why? Under URSA's metric-dependent corruption kernel, the relationship between tt and retained information is highly nonlinear:

tt0.050.350.650.800.900.950.99
P(xt=x^)P(x_t = \hat{x})0.003%0.04%0.73%4.2%20%53%97%

For t≤0.65t \leq 0.65, fewer than 1% of the student's tokens survive corruption, encouraging the distribution matching objective to focus on global clip reconstruction. In contrast, at large tt, particularly t≥0.95t \geq 0.95, much of the student's generated structure is preserved, and distribution matching primarily evaluates local token predictions conditioned on that structure.

Our hypothesis is that the low-noise signal provides little benefit for a one-step generator. Since the corruption time only affects distribution matching (the critic never observes xtx_t), including the full range slow the sharpening of student predictions. This alone does not cause instability. However, when the critic is active from the first step, adversarial gradients are easy to dominate before the student learns to commit to discrete tokens. With t∈[0.05,0.65]t \in [0.05, 0.65], the student is able to sharpen its predictions quickly, allowing the two objectives to reach a stable balance.

A practical consequence is that, with the restricted interval, the distribution-matching and adversarial losses can be trained jointly from the first update without loss in quality: we trained this way and reaches 82.25–82.57, against 81.78 for the staged schedule of the submission. In contrast, our joint-training run using the full range diverged. Notably, Soft-Di[M]O, which samples tt over the full range, also introduces the adversarial generator loss only after a distribution-matching warm-up.

Q2. Can you elaborate on the necessity of the augmentation and of sharing it across teacher and student latents? Since the mask is binary, each token in the real latent always belongs to the codebook while the student's latent can be an interpolation. Even with such augmentation, the shortcut still exists. Why is this not an issue?

A3. Thank you for highlighting this important observation. The real latent consists of codebook vectors, whereas the student latent consists of expected embeddings. Binary replacement does not eliminate this distinction. We acknowledge that our original statement that shared replacement "prevents a shortcut" was inaccurate, and we have removed this claim in the updated manuscript. Below, we clarify two points: (1) our results do not rely on this shortcut, and (2) the augmentation is not essential to our method.

(1) The results do not rely on the shortcut. App. H.5 provides a direct ablation. With the straight-through interface, the critic receives only codebook vectors for both real and generated samples, eliminating the "vertex vs. interior" cue by construction. This variant achieves 81.55 on VBench, compared with 81.68 using the expected embeddings in our paper and 81.59 when the expected embeddings are rescaled to match the norm of codebook vectors. These comparable results suggest that the shortcut has little impact on the final performance.

Why we use the augmentation. This is an inherited component. Corrupting discriminator inputs for regularization dates back to Diffusion-GAN and was subsequently adopted for distillation in DMD2. Soft-Di[M]O extended this approach to discrete models by replacing a random subset of real token embeddings and generated soft embeddings with the mask embedding, using the same mask for both samples (Sec. 3.2 and App. G.3). We follow this design, with one modification required by URSA: since URSA has no mask token, replaced positions receive random codebook entries instead.

(2) Is the augmentation necessary? We conducted 100-step ablation forks and evaluated VBench Total after 50 and 100 updates:

Replacement before the criticnoneindependent drawsshared draw (ours)
VBench Total80.58 / 81.5480.88 / 81.4380.11 / 80.95

The augmentation provides no measurable benefit in these experiments. These results suggest that neither the augmentation nor sharing the replacement is necessary for DDMD-V. We therefore explicitly clarify that it is optional in the revised manuscript and regard it as a regularization operation.

Reviewer_TsUPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DDMD-V, which distills a uniform discrete video diffusion model into a one-step generator. Student videos are corrupted using the teacher’s native probability path, and a teacher–auxiliary categorical divergence supplies a surrogate update to student logits. A spatiotemporal discriminator operating on expected FSQ embeddings provides additional supervision over multi-token video structure.

Critical strengths
  • The performance is good to me. The student generates 49-frame, 320×512 videos in one forward pass, obtaining 81.78 VBench compared with 81.75 for the 50-step teacher.

  • The authors conduct comprehensive evaluations. Teacher sampling-budget sweeps, VideoReward, and human preference comparisons support the result. Ablations examine temporal critic structure, corruption coverage, divergence direction, and hard/soft critic inputs.

Critical weaknesses
  • The method novelty should be positioned precisely. Expected embeddings and GAN-based refinement already appear in Soft-Di[M]O (https://arxiv.org/abs/2509.22925). The contribution here is their adaptation to uniform discrete video diffusion, including the native corruption path and spatiotemporal critic. A controlled adaptation baseline would better isolate these video-specific improvements.

  • Aggregate parity does not imply uniform preservation. Relative to the teacher, the student’s Semantic score decreases from 76.88 to 76.11, Aesthetic Quality from 62.10 to 60.53, and Imaging Quality from 62.23 to 59.10. These trade-offs deserve more explicit discussion.

  • Generality and distribution coverage remain unclear. Experiments use one teacher, tokenizer, resolution, and clip length.

Questions for authors

See weaknesses.

2 comments

We thank the reviewer for the positive assessment. We address each concern below.

W1. The method novelty should be positioned precisely. A controlled adaptation baseline would better isolate these video-specific improvements.

A1. DDMD-V is the first method that distills a uniform discrete diffusion video model into a one-step generator that matches its 50-step teacher on the VBench aggregate. We would also like to emphasize that adapting previous one-step image distillation methods to uniform discrete video is non-trivial. Without spatiotemporal critic, the student can not achieve performance comparable to that of the teacher. Without the corruption-time restriction, distillation fails when the distribution-matching and adversarial losses are applied jointly from the first update.

We implemented Di[M]O, Soft-Di[M]O, DDMD-V without the corruption-time restriction on URSA and trained them under the same setting. The performance is listed below:

MethodVBench Total
Teacher (1 step)58.56
Di[M]O-style74.59
Soft-Di[M]O-style78.69
DDMD-V w/ full t range52.59
DDMD-V w/ t ∈ [0.05, 0.65]81.78
Teacher (50 steps)81.75

W2. Aggregate parity does not imply uniform preservation. Semantic, Aesthetic Quality and Imaging Quality decrease. These trade-offs deserve more explicit discussion.

A2. To determine whether these differences are specific to a particular checkpoint or to the one-step student itself, we retrained DDMD-V and evaluated every checkpoint saved during the final 500 training updates.

ModelQualitySemanticSubjectActionObjectColorAestheticImaging
Teacher, 16 steps82.2575.1795.0094.4092.2377.2261.0958.80
Teacher, 32 steps82.8976.9495.6995.6093.8079.8561.9561.12
Teacher, 50 steps82.9676.8895.7993.6092.6679.8062.1062.23
Ours, submitted model83.2076.1096.2397.0092.5683.7160.5359.10
Ours, update 1,55081.2275.6697.4997.8093.8179.7561.3460.54
Ours, update 1,60082.3275.4997.1097.0092.4881.5460.7259.83
Ours, update 1,65082.6475.2096.9996.8091.8882.6960.6358.39
Ours, update 1,70083.1176.6596.8697.4093.1685.7160.0359.49
Ours, update 1,75083.1476.3496.4497.6092.1783.2860.8160.00
Ours, update 1,80082.9675.1695.1697.2092.0383.1660.4859.28
Ours, update 1,85083.6876.5696.3997.8092.2683.4560.9360.19
Ours, update 1,90082.1675.7797.0997.2091.4781.2061.0060.32
Ours, update 1,95082.4676.3996.8496.2092.9481.9260.9759.69
Ours, update 2,00082.5075.4496.1396.4092.1081.3060.1359.48
Checkpoints above the 50-step teacher3 / 100 / 109 / 1010 / 103 / 109 / 100 / 100 / 10

The pattern is consistent across checkpoints: Subject, Action and Color are above the 50-step teacher at most checkpoints, while Semantic, Aesthetic Quality and Imaging Quality are below. Notably, Semantic and Imaging Quality are between the level of 16- and 32-step teachers.

Why per-frame fidelity is lower. The teacher's later sampling steps primarily refine fine-grained visual details. Increasing the number of steps from 16 to 50 improves the teacher's Imaging Quality by 3.43 and Aesthetic Quality by 1.01, while the Total score increases by only 0.95. A single step sacrifices part of this.

Why subject, action and color are higher. The spatiotemporal critic supervises multi-frame structure directly: without it, Subject Consistency falls from 95.77 to 88.53 (App. H.1).

W3. Generality and distribution coverage remain unclear. Experiments use one teacher, tokenizer, resolution, and clip length.

A3. This is a fair limitation and we now state it explicitly. To our knowledge, URSA is the only publicly available family of uniform discrete diffusion models for video generation, and its released video checkpoints share the same tokenizer, resolution and clip length, so these factors cannot be varied at present.

What can be varied is the teacher. We applied the same training recipe on URSA-0.6B:

TeacherTeacher, 50 stepsTeacher, 1 stepDDMD-V, 1 step
URSA-1.7B (paper)81.7558.5681.78
URSA-0.6B (new)80.9257.3581.14

Similarly, student model can achieve comparable VBench with one single pass.

Moreover, we'd like to emphasize that the DDMD-V objective relies only on the teacher's forward corruption kernel and its clean-token posterior, both of which are defined for any uniform discrete diffusion model. Nothing is specific to URSA's architecture or tokenizer. We hope this can address the concern.

Thank you for the response.

The table at A1 shows that DDMD-V achieves 81.78 compared to Di[M]O-style's 74.59 and Soft-Di[M]O-style's 78.69, which supports that adapting the previous one-step distillation techniques to uniform discrete diffusion models is not trivial. This addresses my first concern.

For W2, I mean though the overall score is good, several subitems are lower than teachers. Considering DDMD-V consumes only one step compared to the teacher's 50 steps and the difference of the worsen metrics is narrow, this performance difference is acceptable. And I also thank the authors for providing the stability across multiple checkpoints.

I also appreciate the acknowledging the generalization limitation since there's only one publicly available video uniform discrete diffusion model.

Given the above reasons, I keep my weak acceptance recommendation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.