- 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
-
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.
-
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
-
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.
-
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
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.
-
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.
-
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 . 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.
-
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 size | 1,600 | 3,200 | 9,600 |
|---|---|---|---|
| VBench Total | 80.18 / 80.85 | 80.12 / 80.80 | 80.11 / 80.95 |
| real-teacher : student fields | 1 : 4 | 1 : 2 | 1 : 1 |
|---|---|---|---|
| VBench Total | 80.31 / 81.07 | 80.11 / 80.95 | 80.10 / 80.60 |
The results indicate both the cache size and the sampling-ratio are not sensitive to the final quality.
- 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.
We thank the reviewer for positive assessment. We address each concern below.
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:
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.
A2. We implemented Di[M]O, Soft-Di[M]O and trained them under the same setting. The results are shown below:
A3. VBench. We performed paired bootstrap resampling over the 946 prompts (B = 2,000):
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):
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.
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.