- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
DistillAlign studies distribution mismatch and mode collapse in multi-stage autoregressive video distillation. It introduces a teacher-normalized evaluation protocol using V-JEPA2 features to measure precision, coverage, and diversity. Building on the Causal Forcing pipeline, it aligns the initialization data with the target DMD teacher and replaces pure DMD refinement with a joint DMD-CD objective, where DMD improves fidelity and CD preserves distribution coverage. Experiments on Wan2.1-T2V show improved VBench scores, coverage, and diversity. Controlled teacher-source swaps, loss-weight ablations, and training-dynamics analyses support the effectiveness of distribution alignment and joint distillation.
Critical strengths
- Clear motivation. The paper clearly identifies two issues in autoregressive video distillation: distribution mismatch between initialization and DMD refinement, and late-stage coverage degradation caused by the mode-seeking behavior of DMD.
- Useful distributional analysis. The teacher-normalized evaluation protocol, together with V-JEPA2 features, precision, coverage, and diversity metrics, provides information that is complementary to VBench and helps analyze distribution coverage during distillation.
- Simple objective with competitive results. The proposed joint objective is conceptually simple. In the reported setting, the weak variant with a 1.3B DMD teacher outperforms the listed 14B-teacher distilled baselines in total VBench and coverage.
Critical weaknesses
- The paper lacks comparisons of training time and GPU memory across different methods.
- The generalization is limited, as experiments are conducted only on Wan2.1-T2V.
- Measuring distribution coverage in the V-JEPA2 feature space may introduce bias, and the robustness of the metrics under different settings is unclear.
Questions for authors
- Will combining two losses significantly increase the time and memory overhead of distillation?
- Does the method generalize to other architectures, such as Wan2.1-I2V or CogVideoX?
7 comments
Official Comment by Authors
mpmn-Q2: Does the method generalize to other architectures, such as Wan2.1-I2V or CogVideoX?
Yes. We evaluated CogVideoX1.5-5B and Wan2.2-TI2V-5B. Starting from a Causal Forcing baseline trained on OpenVid, we progressively add teacher-source alignment and joint distillation. For each architecture, both the DMD teacher and the reference teacher for distribution metrics are its corresponding base model.
The results are reported below.
| Model / Setting | VBench | Distribution | |||
|---|---|---|---|---|---|
| Total | Quality | Semantic | Precision | Coverage | |
| CogVideoX1.5-5B (base model) | 82.67 | 83.41 | 79.71 | / | / |
| baseline | 82.32 | 83.11 | 79.16 | 0.68 | 0.43 |
| + Teacher-source alignment | 82.41 | 83.18 | 79.33 | 0.75 | 0.52 |
| + joint distillation | 82.46 | 83.22 | 79.42 | 0.79 | 0.59 |
| Wan2.2-TI2V-5B (base model) | 83.19 | 84.53 | 77.83 | / | / |
| baseline | 83.44 | 84.87 | 77.72 | 0.81 | 0.49 |
| + Teacher-source alignment | 83.81 | 85.23 | 78.13 | 0.74 | 0.63 |
| + joint distillation | 83.90 | 85.31 | 78.26 | 0.76 | 0.66 |
The gains in VBench scores and coverage support the applicability of our method to both architectures.
mpmn-W1: How reliable is V-JEPA2-based coverage, and how robust are the conclusions to evaluation settings?
[Accuracy]
Coverage depends on relative distances in V-JEPA2 feature space. To validate this ordering, we plan a study of 50 anonymized comparisons, each containing a reference video and two candidates with a clear preference. Thirty raters would assess 10 questions each, yielding six independent ratings per question. We would compare feature-distance preferences with human preferences.
| Agreement on clear-preference pair | Agreement rate |
|---|---|
| 46/50 | 92.00% |
[Robustness]
Using the DMD stage as an example, we fix the initialization checkpoint and vary the training seed. The table below reports means and standard deviations.
| Initial coverage (%) | After-DMD coverage (%) | Δ coverage (%) | Δ precision (%) | Δ diversity (%) |
|---|---|---|---|---|
| 61.33 | 57.03 ± 1.66 | 3.83 ± 0.75 | 0.31 ± 0.64 | 1.16 ± 0.58 |
The standard deviations of the changes in coverage, precision, and diversity are all below 1%.
Seeking Further Feedback
Thank you again for your thoughtful review and constructive feedback. As the discussion phase is coming to a close, we would greatly appreciate any further thoughts on our rebuttal. If any concerns remain or additional clarification would be helpful, please feel free to let us know, and we would be happy to address them. If the main concerns have been sufficiently addressed, we would sincerely appreciate your consideration of updating your assessment accordingly.
Thank you again for your time and consideration.
Official Comment by Authors
Thank you for the constructive review. We organize your comments into individual questions and respond to them in order below.
[Training time and computation]
In the default DMD stage, the student is updated once per five critic updates, and joint distillation adds computation only to student updates. We report individual updates and a full cycle (five critic updates plus one student update) below. FLOPs are reported per GPU. Measurements were obtained on a single node with eight A100 GPUs.
Full-cycle time increases by 8.24% and FLOPs by 6.37%, which we consider acceptable overhead.
[Memory]
We report peak allocated GPU memory, time-averaged device memory usage, and peak device memory usage below.
Peak device-memory usage is 69.08 GiB, within the memory capacity of an 80-GB A100.