- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes Radian, a post-training framework that augments on-policy distribution matching distillation with representation-space adversarial supervision for autoregressive video generation. Generated rollouts are sparsely decoded into RGB frames and processed by a frozen visual encoder, with lightweight discriminator heads trained against real video frames. The framework uses ODE initialization, discriminator calibration, and joint DMD-adversarial refinement, without changing the generator's inference architecture. Experiments cover four-step chunk-wise generation, one-step frame-wise generation with first-frame enhancement, and minute-long rolling generation on Wan2.1-1.3B. The paper also compares image, video, and diffusion-internal discriminator representations, reporting improvements on VBench, VideoAlign, and VBench-Long.
Critical strengths
-
The method provides a practical way to incorporate real-data perceptual supervision into an autoregressive distilled generator without increasing inference cost. Combining a teacher-based objective with adversarial feedback on decoded outputs is well motivated, and the reported results suggest that this combination can improve perceptual quality under small denoising budgets.
-
The empirical evaluation covers several generation regimes and a useful range of discriminator representations. Comparisons across image encoders, video encoders, diffusion features, encoder sizes, and feature depths provide evidence about practical design choices. The use of matched evaluation prompts and 15 inference seeds, together with qualitative comparisons, strengthens the empirical investigation.
Critical weaknesses
-
The methodological novelty beyond existing adversarial distillation is somehow limited. ADD already combines teacher distillation with adversarial supervision through frozen visual features and lightweight multi-level discriminator heads. Radian primarily adapts this recipe to on-policy autoregressive video distillation, with sparse decoding and staged optimization. While this adaptation has practical value, the paper does not sufficiently establish which video-specific design choices are essential or what additional insight distinguishes the method.
-
The method exhibits substantial sensitivity to the external encoder and feature configuration, limiting confidence in the robustness of the proposed recipe. In Table 2, replacing DINOv2 with V-JEPA 2.1 reduces Dynamic Degree from 0.8056 to 0.5991 and VideoAlign Total from 0.8033 to 0.5892. Tables 3 and 4 also show considerable variation across encoder sizes and feature-layer selections.
Questions for authors
-
Table 2 shows that replacing DINOv2 with V-JEPA 2.1 substantially reduces Dynamic Degree from 0.8056 to 0.5991, while VideoAlign Motion Quality also decreases from 0.2094 to 0.1275. Why does a video encoder with explicit temporal modeling produce this behavior compared with the image-based DINOv2 encoder? Does this reflect a bias toward low-motion outputs, or differences in input sampling, discriminator design, and effective adversarial gradient strength? Does the gap persist after separately calibrating the adversarial weight and optimization settings?
-
DMD can exhibit mode-seeking behavior which would cause motion degradation, while the default frame-wise adversarial objective does not explicitly constrain temporal dynamics. This raises the concern that prolonged joint training could improve individual-frame quality at the expense of motion diversity or action progression. Does Radian exhibit motion collapse or other quality regressions when trained beyond the reported schedule?
I'm willing to raise my score if the authors satisfactorily address these concerns, particularly the sensitivity to encoder choice and the stability of motion quality and diversity under prolonged training.
5 comments
Official Comment by Authors (2/3)
| Config. | GAN W. | D LR | G LR | Aggregation | VBench Total | Dynamic | Aesthetic | Imaging | VideoAlign Total | VQ | MQ | TA | GNR. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Ours | 0.02 | 1e-5 | 2e-6 | Multi-head discriminator | 0.8356 | 0.5991 | 0.6823 | 0.7031 | 0.5937 | 0.0506 | 0.1273 | 0.4119 | 1.5856 |
| GAN W. | 0.1 | 1e-5 | 2e-6 | Multi-head discriminator | 0.8337 | 0.5658 | 0.6813 | 0.6902 | 0.5748 | 0.0452 | 0.1187 | 0.4109 | 2.7429 |
| GAN W. | 0.5 | 1e-5 | 2e-6 | Multi-head discriminator | 0.8268 | 0.4706 | 0.6783 | 0.6575 | 0.5366 | 0.0353 | 0.1057 | 0.3956 | 7.8147 |
| D LR | 0.02 | 5e-6 | 2e-6 | Multi-head discriminator | 0.8349 | 0.5757 | 0.6804 | 0.7103 | 0.5675 | 0.0435 | 0.1194 | 0.4046 | 1.0361 |
| D LR | 0.02 | 3e-5 | 2e-6 | Multi-head discriminator | 0.8312 | 0.5403 | 0.6758 | 0.6902 | 0.5568 | 0.0404 | 0.1106 | 0.4058 | 1.6608 |
| G LR | 0.02 | 1e-5 | 5e-7 | Multi-head discriminator | 0.8278 | 0.5806 | 0.6793 | 0.6995 | 0.5611 | 0.0436 | 0.1203 | 0.3972 | 1.1392 |
| G LR | 0.02 | 1e-5 | 5e-6 | Multi-head discriminator | 0.8284 | 0.4908 | 0.6832 | 0.7004 | 0.5120 | 0.0357 | 0.0856 | 0.3907 | 1.6594 |
| G LR | 0.02 | 1e-5 | 1e-5 | Multi-head discriminator | 0.8029 | 0.3509 | 0.6653 | 0.6806 | 0.4064 | 0.0103 | 0.0457 | 0.3504 | 1.8748 |
| Aggregation | 0.02 | 1e-5 | 2e-6 | Spatial + Temporal | 0.8141 | 0.3877 | 0.6702 | 0.6853 | 0.4206 | 0.0297 | 0.1106 | 0.2803 | 1.3529 |
| Aggregation | 0.02 | 1e-5 | 2e-6 | Q-Former | 0.8052 | 0.3207 | 0.6623 | 0.6775 | 0.4600 | 0.0153 | 0.1009 | 0.3438 | 1.4438 |
Aggregation designs: Multi-head discriminator = Multi-Discriminator Heads; Spatial + Temporal = Spatial Average + Temporal Conv + Multi-Discriminator Heads; Q-Former = Q-Forme Heads.
Metrics: GAN W. = Weight of the GAN Branch; GNR. = Weighted GAN/DMD Gradient-Norm Ratio at Step .
(4) Sensitivity to Frame Sampling Rate: We further investigate the influence of temporal sampling by varying the number of discriminator input frames and sampling stride. The results show that both frame count and temporal stride affect generation quality and dynamics. Our default configuration ( frames, stride ) achieves the best overall VBench and VideoAlign scores, while other configurations exhibit different trade-offs between motion magnitude and perceptual quality. These observations highlight the importance of temporal sampling design for effective adversarial supervision.
| Configuration | VBench Total | Dynamic | Aesthetic | VideoAlign Total | MQ | TA |
|---|---|---|---|---|---|---|
| 16 frames, stride 2 | 0.8319 | 0.5558 | 0.6804 | 0.5576 | 0.1073 | 0.4051 |
| 32 frames, stride 1 | 0.8332 | 0.5729 | 0.6811 | 0.5761 | 0.1178 | 0.4149 |
| 32 frames, stride 2 (Ours) | 0.8356 | 0.5991 | 0.6823 | 0.5937 | 0.1273 | 0.4119 |
| 32 frames, stride 4 | 0.8302 | 0.5325 | 0.6784 | 0.5601 | 0.1105 | 0.3975 |
| 64 frames, stride 2 | 0.8328 | 0.6055 | 0.6707 | 0.5842 | 0.1203 | 0.4137 |
Overall, DINOv2 achieves better overall performance under the configurations selected in our paper. V-JEPA's Motion Quality is affected by hyperparameters and architecture choices, and the performance gap may also depend on supervision strength, input sampling, discriminator design and frame sampling strategy. Nevertheless, V-JEPA generally performs worse across multiple configurations, suggesting that its representations may exhibit an inductive bias toward low-motion outputs. Its discriminator heads may more readily reward consistent, conservative generations without equally constraining motion diversity and naturalness, thereby encouraging the generator to reduce motion. We therefore retain the current configuration in our paper. On the other hand, experiments on an internal closed-source TAI2V model show that V-JEPA-GAN can improve speech-prompt following more effectively than DINO-GAN in certain digital-human scenarios. This is also partly reflected in its VideoAlign Text Alignment score of , suggesting that V-JEPA-GAN remains useful in specific applications.
Official Comment by Authors (3/3)
Q3: Prolonged Training and Motion Degradation
We explicitly acknowledge that the mode-seeking behavior associated with DMD can also occur in our method, particularly during prolonged training, where noticeable motion degradation may emerge.
However, our extended training results show that motion degradation does not occur immediately. In fact, both dynamic degree and motion quality improve during the earlier stages of training, with noticeable degradation emerging only after steps, around step . During Stage III training of DINO-GAN, we randomly tracked checkpoints from to steps across seeds and averaged the results. Dynamic Degree increases from at step to at step and at step , before dropping to at step and continuing to decline. Meanwhile, VideoAlign Motion Quality increases from at step to at step , then decreases to at step . Thus, we observe no substantial motion collapse during the earlier training stages.
This issue also exists in video-VFM GANs and may emerge earlier and more severely. In our V-JEPA joint training experiments, these discriminators tend to favor consistent, conservative outputs, making the generator more prone to reduced motion. Dynamic Degree decreases from at step to at step . Meanwhile, VBench Total remains nearly unchanged, even increasing slightly from to . Therefore, introducing a video feature extractor that explicitly models temporal relationships is not sufficient to prevent motion degradation, and aggregate scores may conceal a substantial decline in motion quality.
In summary, our method cannot fundamentally eliminate the inherent limitations of DMD, and we will explicitly acknowledge this limitation in the revised paper. We do not claim to completely resolve DMD's mode-seeking problem. Rather, while mode seeking remains present, our goal is to cover more diverse, higher-quality generation modes and improve perceptually sensitive dimensions such as human preference, color, subject consistency, and motion.
Official Comment by Authors (1/3)
Thank you for your comments. Here are our responses to your comments:
Q1: Novelty and Video-Specific Design
We thank the reviewer for this comment and agree that frozen visual features and lightweight multi-level discriminator heads are not entirely new. Our implementation draws on StyleGAN-T and shares similarities with ADD. More precisely, our contribution lies in adapting this approach to autoregressive video distillation through on-policy conditioning, sparse decoding, and staged optimization. In addition, our pair-rank mechanism distributes the generator and VAE decoder/VFM/discriminator heads across two GPUs, forming one logical training unit through cross-device gradient propagation to alleviate the memory pressure of long-video training. Our method also enables video VFMs to extract features from an entire clip while preserving the computational graph, allowing GAN gradients to train the generator through clip-level supervision. We will revise our novelty claims and clarify the existing limitations.
Q2: V-JEPA 2.1 Encoder and Hyperparameter Sensitivity
Regarding whether the performance gap between DINOv2 and V-JEPA 2.1 arises from the encoders themselves or mismatched hyperparameters, we conducted the following analyses. All hyperparameter and architecture comparisons were evaluated using checkpoints trained for 800 steps.
(1) Different Training Stages: Under the configuration retained in our paper, V-JEPA's Dynamic Degree decreases from 0.6269 at step 200 to 0.6108 at step 500 and 0.5991 at step 800, remaining within a relatively low range.
(2) Hyperparameter and Architecture Calibration: We tested different GAN weights (1.0, 0.5, 0.1, 0.02), discriminator learning rates (5e−6, 1e−5, 3e−5), generator learning rates (5e−7, 2e−6, 5e−6, 1e−5), and three feature aggregation designs: multi-level discriminator heads, spatial average pooling followed by temporal convolutions and multi-level heads, and a Q-Former global head that summarizes the entire video through learnable queries. Other settings, including the G:D update frequency (1:5), batch size (8), and training duration (800 steps), were kept consistent. We ultimately selected a GAN weight of 0.02, D LR of 1e−5, G LR of 2e−6, and multi-level discriminator heads. Detailed results are shown in the table below.
(3) Gradient Strength: V-JEPA training logs show that the weighted GAN/DMD gradient-norm ratio increases from 0.2432 at step 200 to 1.5856 at step 800, indicating that the effective adversarial supervision strength changes throughout training. The corresponding ratio for DINO-GAN is approximately 1.7984, suggesting that V-JEPA-GAN's gradient strength remains within a reasonable range.