acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Aware Consistency Distillation for Few-Step Video Generation

Abstract

We study few-step video generation, i.e., distilling a multi-step video generator, which typically requires tens of sampling steps, incurring substantial latency and compute, into a few-step student. Consistency distillation is a common recipe, in which a multi-step teacher provides the consistency targets for a few-step student. However, these teacher-guided targets are not equally trustworthy, and the content is harder to learn where it varies rapidly over time, e.g., moving foliage shadows or flowing water. We observe that supervision reliability follows the local difficulty of the content rather than semantic complexity: regions that change little yield consistent endpoint predictions, whereas regions with large temporal variation produce larger discrepancies that coincide with the largest perceptual errors. Motivated by this observation, we propose **Uncertainty-Aware Consistency Distillation** (UACD), which reweights consistency supervision at each spatiotemporal region using a local, parameter-free uncertainty estimate. Specifically, we construct two independently perturbed teacher-guided consistency paths, whose student endpoint predictions provide a consensus target; the discrepancy between the student's direct prediction and this target is the uncertainty proxy. We then relax the consistency penalty on high-uncertainty regions through an exponential weight, while keeping the full penalty elsewhere, since the student cannot be expected to match targets that are hard to learn. To preserve perceptual quality under aggressive step reduction, we integrate feature-space adversarial training with semantic alignment. With parameter-efficient LoRA adaptation of the 50-step Wan model, our method achieves state-of-the-art 4-step generation on VBench 2.0 (0.556 mean score) and is preferred over competing methods in a user study.

Primary area:
generative models
Keywords:
Few-Step Video Generation, Consistency Distillation, Uncertainty-Aware Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yBLKOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes UACD, which averages student endpoint predictions from two independently perturbed, teacher-advanced paths to construct a detached target. The squared discrepancy between a direct student prediction and this target determines an exponential weight on the consistency loss. Feature-space adversarial training and discriminator semantic alignment complement the uncertainty objective. On Wan2.1-1.3B, the method achieves 0.556 on VBench2.0, compared with 0.527 for DCM and 0.452 for its 50-step teacher. Ablations investigate uncertainty estimation, path count, and semantic alignment.

Critical strengths
  • This paper proposes a simple, explicitly specified adaptive supervision paradigm. The uncertainty proxy requires no additional learned network, keeping the central intervention compact. The element-wise weighting preserves fine spatial and temporal granularity, making it possible to attenuate selected latent locations rather than discard entire training examples.
  • The paper conducts comprehensive component studies and provides enough implementation information. The ablations include single-path estimation, a learnable alternative, and removal of semantic alignment, helping assess several design choices. Supplementary experiments vary uncertainty strength, alignment strength, and feature depth, exposing sensitivity beyond the main configuration.
Critical weaknesses
  • Concerns about baseline reproduction and comparison validity. On the provided webpage, the qualitative comparisons show conspicuous degradation in some baseline outputs, particularly the unnatural colors in DCM and severe oversaturation in Causal Forcing. These artifacts raise concerns about whether the baselines were evaluated with appropriate checkpoints and method-specific inference settings. The authors can specify the official code versions, sampling schedules, guidance settings, and decoding configurations, and validate the recommended setups of the baselines. If the same generation pipelines were used for Table 1 and the user study, any reproduction issues could affect both the quantitative and preference conclusions.
  • Insufficient coverage of directly comparable baselines. Table 1 includes only one directly comparable bidirectional few-step distillation baseline, DCM, while other distilled baselines are autoregressive. Closely related baselines, particularly rCM, DMD2, DOLLAR, Self Forcing, and CausVid, should be evaluated under a matched backbone and sampling budgets. Moreover, total NFE and full-video latency do not capture the streaming benefits of autoregressive generation. I recommend separating the two categories, specifying student sizes and generation configurations, and doing the comparison under the same settings.
Questions for authors
  • The uncertainty is insufficiently justified. Eq. 3 measures direct to consensus residual, there exists a case that U=0 despite potentially large path disagreement. Thus the interpretation linking two paths to higher U is not guaranteed. Besides, a large residual may arise from an inaccurate student even when the target is reliable. So the definition alone cannot identify which side is erroneous?
6 comments
AuthorsOpenReview ↗

Response to W1: Baseline Reproduction and Evaluation Configurations

We thank the reviewer for raising concerns about the validity of our baseline comparisons. We would like to clarify that we use the official checkpoints and inference implementations released by the respective methods, following their recommended sampling configurations without additional tuning. All methods are decoded with their own official VAE, with no post-processing, no re-encoding and no frame interpolation. The same generation pipelines were used for Table 1 and the user study.

Specifically, DCM uses the official DCM_WAN checkpoint with four sampling steps at the official timesteps [1000, 982, 949, 857], no CFG. We also note that some videos on DCM's official project page exhibit similar unnatural color variations. Causal Forcing uses the official Frame-wise 4-step checkpoint with frame-wise generation and the official timesteps [1000, 750, 500, 250], no CFG.

AuthorsOpenReview ↗

Response to W2: Additional Comparisons with Related Baselines

We have conducted additional experiments with rCM, Self Forcing, and CausVid, using their publicly released Wan checkpoints. All methods are decoded with their own official VAE, with no post processing, no re encoding and no frame interpolation.

  • rCM: rCM_Wan2.1_T2V_1.3B_480p.pt
  • Self Forcing (autoregressive model): self_forcing_dmd.pt
  • CausVid (autoregressive model): autoregressive_checkpoint/model.pt

We clarify that DMD2 is an image-generation distillation method rather than a directly applicable video-generation baseline. DOLLAR does not release publicly available code or checkpoints, so we cannot conduct a reproducible evaluation under our experimental setup.

The additional results are summarized below:

MethodStepsTeacherStudentHuman FidelityCreativityControllabilityCommonsensePhysicsMean
Self Forcing4Wan2.1-T2V-14BWan2.1-T2V-1.3B0.8480.4670.2520.5760.4530.519
CausVid4Wan2.1-T2V-1.3BWan2.1-T2V-1.3B0.8380.4340.2440.5330.4280.496
rCM4Wan2.1-T2V-1.3BWan2.1-T2V-1.3B0.8370.5160.2360.5960.4530.528
UACD (Ours)4Wan2.1-T2V-1.3BWan2.1-T2V-1.3B0.8610.6420.2030.5560.5160.556
AuthorsOpenReview ↗

Response to Q1: Interpretation and Limitations of the Uncertainty Proxy

We acknowledge that UU is an imperfect proxy: it may fail to capture inter-path disagreement and cannot distinguish target inaccuracy from student prediction error. Therefore, UU should be interpreted as a measure of local prediction inconsistency, rather than a direct measure of target reliability.

We clarify that UU is a proxy for local consistency instability, defined by the discrepancy between the direct student prediction and the consensus prediction, rather than a measure of inter-path disagreement or calibrated target uncertainty. Its purpose is not to identify incorrect targets or distinguish target ambiguity from student prediction error, but to adaptively modulate the strength of the consistency constraint.

Specifically, a large UU indicates that the direct prediction and the consensus target are poorly aligned at the corresponding location. In this case, enforcing a strict latent-level match to the consensus target may impose an overly restrictive constraint. We therefore reduce the corresponding consistency penalty while retaining adversarial and other auxiliary supervision. We acknowledge that this strategy may also reduce consistency supervision in regions where the student has not yet learned sufficiently. This is an inherent limitation of using prediction discrepancy as a parameter-free proxy, rather than an explicit estimator of target correctness.

We have revised the manuscript to clarify this interpretation and have avoided implying that large UU necessarily indicates large inter-path disagreement or an unreliable target.

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

This paper distills Wan2.1 1.3B video model to video generation model with 4-step sampling. It puts independent noise to one video to make two noisy inputs. The teacher moves each one as one Euler step, and uses average of two endpoints predicted by student as a consistency target. Square of the difference between the student's directly predicted value and the target is uncertainty, and it gives weaker supervision using exponential weight where the value is big. During training it uses adversarial loss from feature space, DINOv2 alignment for discrimiinator, and train LoRA. It achieved positive score on VBench 2.0 and User study.

Critical strengths
  1. I think the comparison with DCM is very convincing. Because two methods both use 1.3B teacher model and 4-step generation. UACD achieved better score on Human Fidelity, Creativity, Physics. For me, this comparison is more meaningful than the comparison with 14B autoregressive model.
  2. Ablation study is also helpful. Table 2 results are as following
  • single path weighted variant : 0.513
  • learnable uncertainty token : 0.509
  • full model : 0.556 Appendix clarifies more about the number of paths and weighting strength. The score is similar even when the paths become three or four (0.554, 0.556), so using two paths is a reasonable choice.
  1. The way to handle hyperparameters is also honest. The average score is higher in lambda = 5 (0.568), but the score could be inflated when lambda is selected in the same benchmark. So the authors fixed the lambda as 1 in advance.
Critical weaknesses
  1. Appendix B explains that the lambda changes the spatial weighting, and also the overall strength of the consistency loss.
  • Table 4 has lambda=0 result. However, this setting returns weight to 1. So it doesn't match the average weight of UACD.
  • Therefore, I'm curious about how much of the performance improvement comes from choosing the region. I'm also curious about how much comes from making consistency term small compared to the adversarial term.
  • In Table 4, the score increases from 0.462 when the lambda=0 to 0.568 when the lambda=5. This trend can also be read in two ways.
  1. It's also unclear that what "U is big" means.
  • U in Equation 3 is a value of square of consistency residual.
  • In Figure 1, U follows the perceptual error. However, the thing that the residual is big can be interpreted in two ways. It might mean that the target is inaccurate and it can also mean that it's where the student model didn't learn yet.
  • The method lowers the supervision in both cases, so we can't know which is right.
  1. Figure 1(c) risk-coverage analysis is useful, but this is a video-level analysis that removes the video where the uncertainty is highest.
  • In the loss, U is divided by its per-video average.
  • So the weight only uses relative difference inside one video, and the video-level size that Figure 1(c) measures is removed.
  • It seems like the connection between local map and local error is mainly shown in heatmap of Figure 1(a), (b).
Questions for authors
  1. I'm curious about the performance if we keep two paths and auxiliary loss as they are, and use uniform consistency weight fit to UACD's average weight.
  2. What is the evidence that big residual means we can't believe the target, rather than the region that student still needs to learn?
  3. Section 4 mentions that the authors used multiple random seeds, but I couldn't find how many seeds were used and variance. Could you report them please?
5 comments
AuthorsOpenReview ↗

Response to Q1: Effect of Spatially Adaptive Uncertainty Weighting

We thank the reviewer for this insightful question. We apologize for the error in Table 4: the previously reported λ=0\lambda=0 results were mistakenly copied from the single-path setting. We have rerun the experiment with the dual-path target construction retained and uncertainty weighting disabled (w=1w=1).To investigate whether the improvement comes from spatially adaptive weighting rather than simply changing the overall consistency loss strength, we conducted two additional experiments while keeping the dual-path target construction and auxiliary losses unchanged: (1) U-mean, which replaces the spatially varying latent-level weights with a uniform weight for each video, computed as the arithmetic mean of its original UACD weight map, and (2) U-shuffle, which preserves the uncertainty weight distribution but randomly shuffles its spatial correspondence.

The results are summarized below:

SettingHuman FidelityCreativityControllabilityCommonsensePhysicsMean
Dual-path, λ=0\lambda=0 (w=1w=1)0.7670.6480.1280.5250.4000.494
Dual-path, U-shuffle0.7560.6370.1310.5160.4630.501
Dual-path, U-mean0.8380.6400.1410.5480.4580.525
Dual-path, UACD (Ours)0.8610.6420.2030.5560.5160.556

Full UACD achieves a mean score of 0.556, outperforming U-mean and U-shuffle by 0.031 and 0.055, respectively. In particular, the improvement over U-shuffle suggests that the benefit does not merely come from the weight distribution itself, but also depends on assigning uncertainty weights to their corresponding spatial tokens. These results suggest that spatial correspondence plays an important role in the effectiveness of uncertainty weighting.

AuthorsOpenReview ↗

Response to Q2: Interpretation of Uncertainty and Risk-Coverage Analysis

We thank the reviewer for raising these important concerns. We clarify that UU represents local consistency instability, rather than a direct measure of target inaccuracy or student prediction error. It reflects the discrepancy between the direct student prediction and the consensus prediction induced by the two teacher-guided paths. Consequently, UU cannot distinguish between target ambiguity and insufficient student learning, nor determine which prediction is incorrect. This is an inherent limitation of using prediction discrepancy as a parameter-free proxy.

A deviation from a particular endpoint does not necessarily indicate an incorrect prediction, as multiple plausible endpoints may exist for a given input. Therefore, we do not treat the consensus target as the uniquely correct solution. Instead, we use UU to adaptively modulate the strength of consistency supervision: a large UU indicates greater local prediction inconsistency, for which enforcing strict consistency matching may be overly restrictive. We thus downweight the corresponding consistency loss while retaining supervision from adversarial and other auxiliary objectives. In this way, high-UU regions are not excluded from learning; only their consistency supervision is reduced.

Figure 1(c) provides complementary evidence of the association between aggregated uncertainty and overall video quality. We perform video-level filtering to examine how the overall quality of the remaining set changes as videos with higher aggregated uncertainty are progressively removed. Figures 1(a) and 1(b), in contrast, provide qualitative visualizations of the spatial variations in our uncertainty proxy within videos.

We have clarified these distinctions in the revised manuscript and avoided interpreting UU as a direct indicator of target reliability or local perceptual error.

AuthorsOpenReview ↗

Response to Q3: Random Seeds and Evaluation Protocol

We thank the reviewer for pointing out the ambiguity in our description. We apologize that the phrase "under multiple random seeds" in Section 4 may have suggested multiple independent evaluation runs. We actually follow the official VBench 2.0 sampling protocol: three videos are generated per prompt for each dimension, except for Diversity, which uses 20 videos per prompt generated with 20 different random seeds. In total, the evaluation covers 1,230 dimension-prompt pairs and 3,860 dimension-level video evaluation instances. Under the deduplicated generation and cross-dimension reuse protocol, these correspond to approximately 3,209 unique source videos. Each dimension is then evaluated and aggregated according to its official scoring protocol.

Therefore, the multiple random seeds refer to repeated video sampling per prompt within the benchmark protocol, rather than multiple independent runs of the entire evaluation. We have clarified this distinction and revised the corresponding description in Section 4.

Reviewer_8CgsOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses quality degradation in few-step video generation. Starting from Wan2.1-1.3B, the method constructs two independently perturbed states from the same training latent, advances each with a frozen teacher, and averages detached student endpoint predictions to form a consistency target. The squared residual between a direct prediction and this target determines a local exponential weight for the consistency loss. Feature-space adversarial training and discriminator alignment to frozen DINOv2 representations complement this objective. The paper reports a four-step VBench 2.0 mean of 0.556, compared with 0.527 for its DCM baseline.

Critical strengths
  1. The problem is practically important. Reducing video-generation cost while retaining motion, visual quality, and prompt alignment is a meaningful objective. The relatively small backbone and parameter-efficient adaptation make the setting relevant to deployment.
  2. The method is explicitly defined and discloses important design choices. The uncertainty weight has a concrete formula and does not require an additional learned uncertainty head. The manuscript specifies stop-gradient treatment and acknowledges in Appendix B that λ changes both relative spatial weights and the overall distillation scale.
  3. The evaluation covers several aspects. The paper includes component ablations, hyperparameter analyses, multiple video-quality dimensions, and a human preference study. The reported gains are potentially valuable and warrant further examination of their source and robustness.
Critical weaknesses
  1. The connection between the dual-path residual and supervision reliability needs stronger justification. Standard consistency distillation constrains different states on the same probability-flow trajectory. Independently noising the same training sample does not guarantee that the resulting generative trajectories have the same endpoint. Consequently, the discrepancy between a direct prediction and a cross-path average can contain legitimate generative variation, rather than only unreliable supervision. For example, let the data and noise be independent standard scalar Gaussians, with x_t=(1−t)x₀+tε. The corresponding exact probability-flow endpoint map is f*(x_t,t)=x_t/√((1−t)²+t²). Even with an exact teacher and student, two independently sampled noises yield a nonzero residual against their averaged endpoint; this discrepancy does not disappear as the teacher integration error tends to zero. This example does not establish that the complete method fails, but it motivates distinguishing legitimate cross-trajectory variation from unreliable supervision. The current risk–coverage analysis does not sufficiently separate these explanations. I do not require a calibrated uncertainty estimator or a general convergence theorem: diagnostics stratified by noise level or motion magnitude, or a suitable same-trajectory control, could also provide convincing evidence for the proxy.

  2. The ablations do not sufficiently separate local weighting from the overall loss-scale effect. Changing λ alters both the relative gradient contributions across locations and the strength of consistency supervision relative to adversarial and feature-matching losses. The paper acknowledges this, but the current results do not establish whether the gains primarily come from spatial reliability modeling or from weakening the overall regression constraint. A uniform-weight control matched to the mean weight, or a control that preserves the weight distribution while shuffling its spatial correspondence, would help distinguish these effects. There is also a specific configuration ambiguity: all six displayed values for λ=0 in Table 4 match Table 2, Variant 2, which does not enable DP. Setting λ to zero does not mathematically remove the dual-path averaged target. Please clarify whether the rows come from the same run and whether the target construction was held fixed in the λ sweep. This is a question about configuration and attribution; identical displayed values alone do not establish their cause.

  3. Evaluation settings that affect the main comparison remain insufficiently specified. The baseline scores appear to involve new VBench 2.0 evaluations, making the actual configurations important. In particular, please specify the DCM four-step checkpoint and schedule, the CogVideoX version and its setup at 832×480 with 81 frames, and whether Causal Forcing uses framewise or chunkwise generation. These choices affect the quality and efficiency comparisons. The per-prompt sample counts and aggregation settings for VBench 2.0 should also be explicit. Human Fidelity covers anatomy, identity, and clothing consistency; it is not equivalent to general frame-level realism. Reporting its three component scores and effective sample counts would make the 0.998 result in Table 2 easier to interpret.

Questions for authors
  1. What evidence distinguishes unreliable supervision in the proposed proxy from legitimate cross-trajectory variation and ordinary student prediction error? Could an analysis stratified by noise level or motion magnitude, or a same-trajectory control, help establish this distinction?
  2. Does the λ=0 experiment retain two independently perturbed paths? With the target construction fixed, how do mean-scale-matched uniform weights or spatially shuffled weights perform?
  3. Can you provide the key baseline checkpoints and sampling configurations, per-dimension sample counts for VBench 2.0, and the three Human Fidelity component scores with their effective denominators?
3 comments
AuthorsOpenReview ↗

Response to Q1: Clarifying the Interpretation of Local Consistency Instability

We thank the reviewer for this insightful comment. As the Gaussian example illustrates, a nonzero residual can arise even with an exact teacher and student. Therefore, the residual cannot be attributed solely to teacher integration error or student prediction error. We thus clarify that UU represents local consistency instability, reflecting the disagreement between the direct student prediction and the consensus prediction induced by the perturbed teacher-guided paths, rather than identifying the specific source of this disagreement.

Our intention is to measure local consistency instability rather than identify the source of disagreement. UU may capture both legitimate cross-trajectory variation and student prediction errors. Importantly, such variations do not necessarily imply substantial semantic differences: predictions from different paths can preserve similar global layouts while differing in local details. Moreover, a discrepancy from a particular endpoint does not necessarily indicate an incorrect prediction, since multiple plausible endpoints may exist for a given input. Therefore, rather than treating the consensus target as the unique correct solution, we use UU to modulate the strength of latent-level consistency supervision. Specifically, large UU indicates stronger local prediction disagreement, for which enforcing a strict match may be overly restrictive. We thus reduce the corresponding consistency loss while retaining supervision from other training objectives.

To further investigate whether the effectiveness of discrepancy-based weighting depends on independent cross-trajectory perturbations, we conducted additional controlled experiments, including a same-trajectory control suggested by the reviewer. The results are summarized below:

SettingUncertainty weightingBranch ABranch BStarting statesVBench 2.0
Original CD + Adv.No----One xtx_t0.462
Single-path UYesDirect Student predictionTeacher 1 step + Student 1 stepSame xtx_t0.513
Dual-path U, Same-trajectory controlYesTeacher 1 step + Student 1 stepTeacher 2 steps + Student 1 stepSame xtx_t0.529
Full UACD (Ours)YesTeacher 1 step + Student 1 stepTeacher 1 step + Student 1 stepIndependently perturbed xt(1),xt(2)x_t^{(1)}, x_t^{(2)}0.556

The additional experiments provide two observations:

  1. Effectiveness without independent perturbations. The single-path variant achieves 0.513, outperforming the original CD baseline (0.462), showing that discrepancy-based weighting can be beneficial without independent perturbations.

  2. Same-trajectory control. The two-branch variant, in which both branches start from the same xtx_t but use different numbers of teacher Euler steps, achieves 0.529, further suggesting that the benefit is not exclusive to independent cross-trajectory perturbations.

Finally, we have revised the manuscript to describe UU as a proxy for local consistency instability, rather than as a direct measure of local supervision reliability. Specifically, large UU reflects stronger disagreement between the direct student prediction and the consensus prediction induced by the two teacher-guided paths, which motivates downweighting the corresponding consistency supervision.

AuthorsOpenReview ↗

Response to Q2: Effect of Global Loss Scaling and Spatially Adaptive Weighting

We thank the reviewer for pointing out this ambiguity. We apologize for the error in Table 4: the previously reported λ=0\lambda=0 results were mistakenly copied from the single-path setting. We have rerun the experiment with the dual-path target construction retained and uncertainty weighting disabled (w=1w=1). Importantly, all four configurations below use the same dual-path target construction, allowing a controlled comparison of different weighting strategies.

We further conducted two experiments to examine the contribution of video-level and spatially adaptive weighting : (1) U-mean, which replaces the spatially varying latent-level weights with a uniform weight for each video, computed as the arithmetic mean of its original UACD weight map, and (2) U-shuffle, which preserves the uncertainty weight distribution but randomly shuffles its spatial correspondence. The results are shown below.

SettingHuman FidelityCreativityControllabilityCommonsensePhysicsMean
Dual-path,λ=0\lambda=0 (w=1w=1)0.7670.6480.1280.5250.4000.494
Dual-path,U-shuffle0.7560.6370.1310.5160.4630.501
Dual-path, U-mean0.8380.6400.1410.5480.4580.525
UACD (Ours)0.8610.6420.2030.5560.5160.556

These results highlight the importance of spatially adaptive uncertainty weighting. Full UACD achieves a mean score of 0.556, outperforming both U-mean (0.525) and U-shuffle (0.501). In particular, the improvement over U-shuffle suggests that the benefit does not come merely from applying a non-uniform weight distribution; rather, it relies on assigning uncertainty weights to their corresponding spatial tokens. This suggests that our uncertainty estimates provide useful spatial information for adaptively modulating consistency supervision.

We have corrected Table 4 and included these additional comparisons in the revised manuscript.

AuthorsOpenReview ↗

Response to Q3: Evaluation Configurations and Sample Counts

We thank the reviewer for requesting further details on the evaluation settings. We use official checkpoints and recommended inference configurations for all baselines. Specifically, DCM uses the official DCM_WAN checkpoint with four sampling steps at timesteps [1000, 982, 949, 857], no CFG. CogVideoX uses the official CogVideoX-5B model through the Diffusers implementation, with 50 inference steps and guidance_scale = 6, generating 81-frame videos at 720×480 resolution. The Diffusers pipeline places no constraint on num_frames at inference, so the full 81 frames are produced with the official weights, unmodified checkpoint and no pipeline patching. 720x480 is the official recommended resolution for CogVideoX-5B; changing it to 832x480 constitutes out-of-distribution inference and would instead lower the baseline performance. All other methods generate 81-frame videos at 832×480 resolution. Causal Forcing uses the official Frame-wise 4-step checkpoint with frame-wise generation and timesteps [1000, 750, 500, 250], no CFG. The remaining baselines also follow their official checkpoints and recommended inference configurations. We have revised the paper to explicitly clarify the resolution difference for CogVideoX-5B.

For VBench 2.0, we follow the official dimension-specific evaluation protocols. The evaluation uses 1,013 unique prompts, with different dimensions using their corresponding prompt subsets. Most dimensions generate three videos per prompt, while Diversity generates 20 videos per prompt. In total, the evaluation covers 1,230 dimension-prompt pairs and 3,860 dimension-level video evaluation instances. Under the deduplicated generation and cross-dimension reuse protocol, these correspond to approximately 3,209 unique source videos. Each dimension is aggregated according to its prescribed scoring and filtering procedure.

For Human Fidelity, we additionally report the three component scores: Human Anatomy, Human Identity, and Human Clothes. The evaluation includes 360 videos for Human Anatomy, 138 videos for Human Identity, and 225 videos for Human Clothes. Note that Human Anatomy retains all evaluated videos, whereas Human Identity and Human Clothes follow their respective validity filtering protocols. We list these components for every compared baseline, and additionally for Variant (1) in Table 2 in the last row. Variant 1 is the ablation configuration, whose mean is the 0.998 reported in Table 2. The results are reported below.

MethodHuman Anatomy (360 videos)Human Identity (138 videos)Identity Valid VideosHuman Clothes (225 videos)Clothes Valid VideosMean (Human Fidelity)
Wan2.1-1.3B0.7170.8061060.6102130.711
CogVideoX-5B0.7490.718970.9252120.797
Causal Forcing0.8340.7901010.9422250.855
OneForcing0.7170.7831070.7622100.754
DCM0.8910.781810.8112120.828
UACD (Ours)0.9150.7681040.9002100.861
Variant (1) in Table 20.9951.00061.0002250.998

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers: Summary of Revisions

We are grateful to all reviewers for the constructive feedback. The questions and weaknesses raised in the reviews have improved the paper substantially.

We address every comment from each reviewer separately below. The revised version of the paper has been uploaded, and it mainly includes the following:

  • adding three further baselines, rCM, SelfForcing and CausVid, together with a table reporting the checkpoint, sampling schedule and guidance setting of every baseline (the videos on the anonymous project page have been updated accordingly).
  • adding two controls that separate the per-video weight from its spatial assignment, namely U-mean and U-shuffle.
  • adding a comparison of three path constructions for UU, from a single path to two independently perturbed paths.
  • unifying the definition of UU across the abstract, the introduction and Section 3.2, so that it is consistently described as a measure of local consistency instability.
  • reporting the three Human Fidelity components for every method, together with the evaluation protocol, and correcting the lambda = 0 row of the weighting ablation.
  • fixing small writing inconsistencies, and extending the limitations to cover the single-run nature of the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.