- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes OPSD-V, a post-training method for few-step autoregressive video generators. A real first chunk initializes the training caches. The student then generates subsequent chunks using its own history. An EMA teacher predicts velocities at the same noisy states, but uses real-video features for an older cached chunk while keeping its latest chunk model-generated. The student matches the teacher at all four denoising steps, starting at chunk 8. Inference retains the original sampler and cache policy.
The authors evaluate the method on Self-Forcing and LongLive using 240 prompts for one-minute generation. They also report ablations and a paired preference study.
Critical strengths
-
The teacher cache offers a useful way to bring real video into training. The student remains on its own rollout, while the teacher can use a cleaner recent history. In the training-free example in Figure 7, replacing the latest generated chunk with real-video KV introduces a visible boundary artifact. Keeping that chunk generated is therefore a reasonable design choice, though its effect after training still needs testing.
-
The method improves both starting models under matched inference settings. OPSD-V raises the reported Quality and Dynamic Degree scores on both backbones. On Self-Forcing, aesthetic and imaging quality also improve. The preference study favors OPSD-V overall, although it covers only ten distinct video pairs per backbone. The denoising-step and latest-KV ablations add useful evidence.
Critical weaknesses
-
The experiments do not yet isolate the benefit of the proposed cache policy. D-OPSD already matches a privileged teacher’s velocities at the student’s own denoising states in text-to-image diffusion. Adapting that idea to a causal video KV cache is worthwhile, but the choice of real and generated history is the key new part here. The cache ablation in Table 2 changes whether the latest teacher-cache chunk comes from the teacher or the student. It does not compare the proposed delayed replacement with an all-real teacher history after training under the same budget. Figure 7 makes a case for the design through a training-free example, but cannot establish which policy produces the better post-trained model. Under the stated attention window, the teacher sees only one recent real-video chunk beyond the shared first chunk and its latest generated chunk, making that comparison particularly relevant.
-
Less degradation late in the video has not been measured directly. The main metrics average over the whole minute. For LongLive, Quality rises from 0.7867 to 0.7992, but its five components other than Dynamic Degree all fall slightly. Dynamic Degree measures how much motion is present, not how good it looks. The human preference results offer some positive evidence, but do not assess the later part of the video separately. The configured full-length training rollout contains 60 chunks, whereas evaluation generates 81. The final 21 chunks, about 16 seconds, therefore extend beyond the training rollout length.
Questions for authors
-
Could you train an all-real-history teacher-cache variant using the same student rollout procedure, velocity loss, data, and training budget? How does its full metric breakdown, especially on later video segments, compare with the proposed delayed replacement?
-
Could you report the existing aesthetic, imaging, subject and background consistency, motion smoothness, and Dynamic Degree scores separately for early, middle, and late segments, with prompt-level uncertainty intervals? In particular, do the final 21 chunks improve on LongLive when Dynamic Degree is reported separately?
5 comments
Response to Reviewer cGMm (2/3) — Q2: LongLive Segment-Level Results
Q2. How does quality change across the rollout and in the final 21 chunks?
We thank the reviewer for suggesting this segment-level analysis. We compare the base models with our OPSD-V checkpoints using the same random seed within each pair. We report six dimensions across the early, middle, and late thirds, as well as chunks 61–81, which extend beyond the 60-chunk training rollout. Each segment is evaluated as a separate video.
From the early third to the final 21 chunks, the mean excluding Dynamic Degree fell by 0.0013 with OPSD-V and 0.0075 without it on LongLive, and by 0.0058 and 0.0132 on Self-Forcing. For both backbones, the paired 95% bootstrap confidence intervals (CIs) for the Imaging quality difference include zero in the early third and lie entirely above zero in the late third and the final 21 chunks.
In the final 21 chunks, the mean excluding Dynamic Degree was higher by +0.00247 (95% CI [−0.00090, +0.00570]) on LongLive and by +0.00587 (95% CI [+0.00238, +0.00937]) on Self-Forcing. Dynamic Degree changed by +0.05342 (95% CI [−0.00112, +0.10876]) and +0.01945 (95% CI [−0.02356, +0.06268]), respectively. On both backbones, OPSD-V had higher Aesthetic quality and Imaging quality point estimates and slightly lower Subject consistency. Background consistency was nearly unchanged.
LongLive
Early · chunks 1–27
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5898 | 0.5880 | −0.00181 | [−0.00632, +0.00272] |
| Imaging quality | 0.7154 | 0.7147 | −0.00070 | [−0.00468, +0.00325] |
| Subject consistency | 0.9362 | 0.9280 | −0.00820 | [−0.01129, −0.00518] |
| Background consistency | 0.9369 | 0.9299 | −0.00699 | [−0.00945, −0.00462] |
| Motion smoothness | 0.9889 | 0.9881 | −0.00079 | [−0.00122, −0.00034] |
| Dynamic Degree | 0.5114 | 0.5946 | +0.08325 | [+0.04428, +0.12458] |
| Mean excluding Dynamic Degree | 0.8334 | 0.8297 | −0.00370 | [−0.00538, −0.00204] |
Middle · chunks 28–54
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5738 | 0.5773 | +0.00356 | [−0.00296, +0.00992] |
| Imaging quality | 0.7059 | 0.7075 | +0.00152 | [−0.00433, +0.00732] |
| Subject consistency | 0.9359 | 0.9320 | −0.00392 | [−0.00901, +0.00166] |
| Background consistency | 0.9425 | 0.9409 | −0.00164 | [−0.00430, +0.00108] |
| Motion smoothness | 0.9869 | 0.9872 | +0.00032 | [−0.00028, +0.00096] |
| Dynamic Degree | 0.5417 | 0.5735 | +0.03188 | [−0.01029, +0.07549] |
| Mean excluding Dynamic Degree | 0.8290 | 0.8290 | −0.00003 | [−0.00288, +0.00289] |
Late · chunks 55–81
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5671 | 0.5720 | +0.00488 | [−0.00219, +0.01184] |
| Imaging quality | 0.7014 | 0.7083 | +0.00688 | [+0.00078, +0.01306] |
| Subject consistency | 0.9341 | 0.9307 | −0.00336 | [−0.00922, +0.00242] |
| Background consistency | 0.9402 | 0.9396 | −0.00056 | [−0.00382, +0.00273] |
| Motion smoothness | 0.9865 | 0.9867 | +0.00023 | [−0.00052, +0.00103] |
| Dynamic Degree | 0.5708 | 0.6237 | +0.05288 | [+0.00919, +0.09860] |
| Mean excluding Dynamic Degree | 0.8259 | 0.8275 | +0.00161 | [−0.00150, +0.00458] |
Final 21 · chunks 61–81
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5661 | 0.5717 | +0.00565 | [−0.00173, +0.01301] |
| Imaging quality | 0.7003 | 0.7080 | +0.00768 | [+0.00075, +0.01456] |
| Subject consistency | 0.9357 | 0.9347 | −0.00100 | [−0.00710, +0.00499] |
| Background consistency | 0.9412 | 0.9410 | −0.00018 | [−0.00339, +0.00292] |
| Motion smoothness | 0.9864 | 0.9866 | +0.00022 | [−0.00059, +0.00105] |
| Dynamic Degree | 0.5812 | 0.6346 | +0.05342 | [−0.00112, +0.10876] |
| Mean excluding Dynamic Degree | 0.8259 | 0.8284 | +0.00247 | [−0.00090, +0.00570] |
Δ is OPSD-V minus Base. We compute the 95% confidence intervals by resampling prompt pairs within each benchmark and weighting the two benchmarks equally. The mean excluding Dynamic Degree gives equal weight to the other five scores.
Response to Reviewer cGMm (3/3) — Q2: Self-Forcing Segment-Level Results and Ethics
Q2 continued. Self-Forcing
Early · chunks 1–27
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5783 | 0.5804 | +0.00209 | [−0.00207, +0.00639] |
| Imaging quality | 0.7181 | 0.7189 | +0.00076 | [−0.00315, +0.00456] |
| Subject consistency | 0.9301 | 0.9240 | −0.00611 | [−0.00905, −0.00314] |
| Background consistency | 0.9372 | 0.9334 | −0.00377 | [−0.00586, −0.00167] |
| Motion smoothness | 0.9862 | 0.9852 | −0.00098 | [−0.00153, −0.00044] |
| Dynamic Degree | 0.6364 | 0.6513 | +0.01491 | [−0.01392, +0.04391] |
| Mean excluding Dynamic Degree | 0.8300 | 0.8284 | −0.00160 | [−0.00326, +0.00013] |
Middle · chunks 28–54
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5505 | 0.5637 | +0.01311 | [+0.00705, +0.01927] |
| Imaging quality | 0.7032 | 0.7124 | +0.00916 | [+0.00268, +0.01577] |
| Subject consistency | 0.9295 | 0.9256 | −0.00388 | [−0.00786, +0.00023] |
| Background consistency | 0.9367 | 0.9344 | −0.00237 | [−0.00507, +0.00031] |
| Motion smoothness | 0.9854 | 0.9848 | −0.00065 | [−0.00134, +0.00007] |
| Dynamic Degree | 0.6397 | 0.6413 | +0.00170 | [−0.03551, +0.03926] |
| Mean excluding Dynamic Degree | 0.8211 | 0.8242 | +0.00307 | [+0.00041, +0.00585] |
Late · chunks 55–81
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5328 | 0.5566 | +0.02380 | [+0.01574, +0.03219] |
| Imaging quality | 0.6988 | 0.7104 | +0.01153 | [+0.00279, +0.02054] |
| Subject consistency | 0.9283 | 0.9229 | −0.00541 | [−0.01052, −0.00052] |
| Background consistency | 0.9320 | 0.9335 | +0.00148 | [−0.00118, +0.00411] |
| Motion smoothness | 0.9849 | 0.9843 | −0.00064 | [−0.00166, +0.00031] |
| Dynamic Degree | 0.6407 | 0.6633 | +0.02264 | [−0.01784, +0.06211] |
| Mean excluding Dynamic Degree | 0.8154 | 0.8215 | +0.00615 | [+0.00295, +0.00950] |
Final 21 · chunks 61–81
| Metric ↑ | Base | OPSD-V | Δ | 95% CI |
|---|---|---|---|---|
| Aesthetic quality | 0.5318 | 0.5561 | +0.02422 | [+0.01569, +0.03291] |
| Imaging quality | 0.6995 | 0.7103 | +0.01073 | [+0.00121, +0.02050] |
| Subject consistency | 0.9328 | 0.9277 | −0.00504 | [−0.01058, −0.00001] |
| Background consistency | 0.9348 | 0.9349 | +0.00014 | [−0.00285, +0.00304] |
| Motion smoothness | 0.9850 | 0.9843 | −0.00072 | [−0.00178, +0.00028] |
| Dynamic Degree | 0.6268 | 0.6463 | +0.01945 | [−0.02356, +0.06268] |
| Mean excluding Dynamic Degree | 0.8168 | 0.8226 | +0.00587 | [+0.00238, +0.00937] |
Δ is OPSD-V minus Base. We compute the 95% confidence intervals by resampling prompt pairs within each benchmark and weighting the two benchmarks equally. The mean excluding Dynamic Degree gives equal weight to the other five scores.
What are the data-use and user-study ethics procedures?
All 3,800 training clips were sourced from YouTube and are licensed under Creative Commons Attribution (CC BY).
We did not seek institutional ethics review before conducting the preference study. All 10 participants consented to participate, and we report only aggregate results.
Response to Reviewer cGMm (1/3) — Q1: All-Real History vs. Delayed Replacement
We invite you to view the videos and tables on our anonymous rebuttal page.
Q1. How does an all-real teacher cache compare with delayed replacement?
We thank the reviewer for suggesting this comparison. We trained All-real variants of LongLive and Self-Forcing. These variants also overwrite the latest completed chunk in the teacher’s KV cache with time-aligned ground-truth KV. This post-training comparison extends our training-free diagnostic in Fig. 7. OPSD-V achieved higher Quality and Dynamic Degree than All-real on both backbones.
All-real lowered full-video Quality from 0.7992 to 0.7179 on LongLive and from 0.8158 to 0.7114 on Self-Forcing. Dynamic Degree showed the largest drop over both the full video and the final 21 chunks. Over the final 21 chunks, Semantic scores rose slightly, while VBench-Long Imaging quality fell from 0.7080 to 0.6193 on LongLive (a decrease of 0.0887) and from 0.7103 to 0.6287 on Self-Forcing (a decrease of 0.0816).
VBench-Long scores
Full-video scores follow the paper’s VBench-Long protocol. Segment scores, including the final-21-chunk table below, evaluate each segment as a separate video.
VBench-Long scores (final 21 chunks)
We evaluate chunks 61–81 as a separate video. Quality is the mean of its six components.
Side-by-side video comparisons