acceptodds
Withdrawn from ICLR 2027

FROM SCORES TO SAMPLES: ELASTIC FORCING FOR AUTOREGRESSIVE VIDEO GENERATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion teacher and an online fake-score model. We instead learn the rollout distribution directly from reference videos, eliminating both score models during post-training. Our framework minimizes maximum mean discrepancy (MMD) in frozen self-supervised video representation spaces, using a hybrid Nystr\"om–Monte Carlo estimator to balance approximation bias and sampling variance. Memory-efficient replay and gradient subsampling make this objective practical. Using the same architecture and initialization as Self-Forcing, our 1.3B model improves the VBench Total score from 83.80 to 84.64 while retaining 17 FPS. Removing auxiliary score models also enables 14B post-training on eight H100 GPUs. Beyond distillation, learning from reference videos enables the acquisition of new visual styles, semantic concepts, and spatial priors without a target-specific diffusion teacher. More demo can be seen on anonymous website: https://video-review-k7m4p9.pages.dev/

Primary area:
generative models
Keywords:
Autoregressive video generation, Diffusion Models, Generative Models, Distribution Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BbpPOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Elastic Forcing, which keeps Self-Forcing's few-step autoregressive rollout but replaces DMD with a squared-MMD objective computed in frozen self-supervised feature spaces. The target distribution is a fixed collection of ~8,000 reference videos, so neither a diffusion teacher nor an online fake-score network is needed during post-training. Two estimators make this practical: a hybrid reference kernel mean that mixes a precomputed Nyström summary of all references with a Monte Carlo minibatch, and "estimation–differentiation decoupling", which evaluates MMD on 256 rollouts but back-propagates through a uniformly sampled subset with Horvitz–Thompson rescaling and segment replay. On Wan2.1-1.3B the VBench Total goes from 83.80 to 84.64. The paper also shows adaptation to monochrome style, a cartoon character and panoramic composition from reference videos alone.

Critical strengths
  • The idea is timely and useful. Removing both score networks is a real systems win and reduce training costs.
  • Decoupling estimation from differentiation is a neat, correct observation: MMD's cross-sample interactions need joint access to features, not joint computation graphs. The unbiasedness and covariance derivations are clear.
  • Learning from references rather than from a teacher opens a genuinely new use: post-training a distilled streaming model toward a target style or concept without first adapting a teacher.
Critical weaknesses
  • The "no teacher" framing is overstated. The 8,000 general-domain references are Wan-generated videos, so the teacher is still used, just offline. The cost of generating them is not included in the training-cost comparison.
  • Gains on the main benchmark are small and unquantified for noise: +0.84 VBench Total for the two-encoder model and +0.45 for the three-encoder model used in all later experiments, from a single training run.
Questions for authors
  1. Which Wan model generated the references? If the 8,000 references come from Wan2.1-14B while Self-Forcing 1.3B distills from Wan2.1-1.3B, the 1.3B comparison mixes the objective change with a stronger implicit teacher. Please state the generator, its sampling steps, and the GPU-hours spent creating the references, and add a Self-Forcing baseline whose teacher matches.
  2. Variance and significance. Please report VBench over several training seeds (or at least several sampling seeds with confidence intervals). Differences such as Nyström-only 84.59 vs hybrid 84.64 are well within what one would expect from noise.
  3. Metric leakage. DINO-family features are used both as a training signal (DINOv3) and inside VBench's subject-consistency metric, and the VLM judge sees only 12–16 frames. Can you add an evaluation that shares no encoder with the training loss (e.g., FVD with I3D, or a larger human study)?
  4. Inconsistencies to resolve:
    • Abstract says 14B training uses "eight H100 GPUs"; Fig. 1 and Sec. 5.1 say H200.
    • Sec. 4 says the 1.3B model trains for 150 steps on 8 H200s; Table 6 says 4 H200s and 100/150 steps.
    • Sec. 5.1 reports a human study with 15 participants; Appendix J describes 42 participants.
    • Table 1/2 give 84.25 VBench Total for three encoders; Table 7 gives 83.81 for the same configuration.
    • Table 8 reports VBench for 14B models although Sec. 5.1 argues VBench is unsuitable at 14B.
    • "Figure ??" in Sec. 3.2; Sec. 3.3 evaluates "Eq. (13)", which lives in the appendix while the main text shows the same loss as Eq. (11); Appendices G and H are duplicates.
2 comments
AuthorsOpenReview ↗

Response to Reviewer BbpP, Part 1

Thank you for recognizing the systems benefit, the selective-gradient analysis, and the flexibility of reference-driven adaptation. We now reply to comments and questions.

[W1, Q1] Reference-generator strength, offline teacher use, and reference preparation costs.

We clarify that Self-Forcing's real-score model is Wan2.1-T2V-14B, whereas our reference videos for the 1.3B comparison were generated by Wan2.1-T2V-1.3B. By all means, we are using a weaker implicit teacher than Self Forcing. The larger target model is therefore on the Self-Forcing side of the comparison, so the gain is not attributable to a larger reference generator.

We agree that “no teacher” was too broad. We mean that no online real/fake score models are used during post-training. We report reference preparation separately from post-training for the following reasons:

  1. Reference generation is an offline cost, and the resulting dataset can be reused across runs.
  2. Elastic Forcing can also leverage publicly available video datasets, such as OpenVid, without generating new references. We use Wan-generated references in our main experiments to reduce the distribution gap from ODE initialization, and report results using OpenVid as reference samples in the Appendix.D.

We agree that the preparation cost should nevertheless be disclosed separately. The requested details are:

Reference preparationDetails
GeneratorWan2.1-T2V-1.3B
Sampling steps50
Generation hardware and GPU-hoursH200 <br/>160 GPU hour
Reference feature / Nyström preparation cost2 GPU hour/0.2 GPU hour

[W2, Q2] Sampling variability and the size of the VBench and estimator gains.

We provide a matched multi-sample evaluation of the closest baseline. All reported evaluations use five samples for each of 946 expanded VBench prompts; the re-evaluation uses the same prompt/seed protocol for both models.

ModelTotal ↑Quality ↑Semantic ↑
Self-Forcing83.9584.7380.83
Elastic Forcing84.6485.4381.48

The improvement is present in all three aggregate metrics. These evaluations average sampling variability, but do not estimate variability across separate training runs. With respect to inference-seed noise, the standard error of the mean VBench score for a given model is 0.0935—far smaller than the 0.69 gap between our model (84.64) and Self-Forcing (83.95). We interpret the benchmark results together with the qualitative and independent evaluations, rather than treating the magnitude of the VBench Total difference alone as decisive.

For Hybrid, adding Monte Carlo comparisons reduces the finite-rank reference bias left by Nyström, while the persistent summary stabilizes the sampled estimate. Its best Total, Semantic, and VLM results in Table 3 support the design. The final Total difference from pure Nyström is modest; the broader contribution also includes video representation selection and estimation–differentiation decoupling.

AuthorsOpenReview ↗

Response to Reviewer BbpP, Part 2

[Q3] DINO overlap with VBench and evaluation using an independent feature extractor.

We agree that evaluation independent of the training encoders is important. Our two-encoder configuration does not use DINOv3 and achieves the highest VBench Total in Table 1, providing a direct check on whether the headline gain depends on DINO supervision.

We additionally evaluate FVD with I3D, comparing each student's outputs against the same evaluation reference collection:

FVD evaluationValue
Reference videos944 × 5 = 4,720 Wan-14B generations
Feature extractorI3D
Self-Forcing FVD ↓266.86
Elastic Forcing FVD ↓217.12
Relative reduction18.6%

The lower FVD supports the quality improvement beyond DINO-based subject consistency. These Wan-14B videos are evaluation references, distinct from the Wan-1.3B training collection. At 14B scale, the blind human study with 42 participants also favors Elastic Forcing (Table 5), providing complementary evidence independent of the training encoders and sparse-frame VLM scoring.

[Q4] Correcting the hardware, update counts, model labels, and cross-references.

We apologize for the reporting errors and clarify the settings below. All post-training experiments use H200 GPUs: the 1.3B model uses four GPUs, with 100 updates for two encoders and 150 for three; the 14B model uses eight GPUs and 80 updates, with five inference steps. The human evaluation involved 42 annotators, not 15.

Table 1 reports the selected endpoints, while Table 7 compares the encoder variants at 100 updates. Table 2 mistakenly copied the mixed-step endpoint results from Table 1; we will correct it to the common-update comparison. This explains the three-encoder scores of 83.81 at 100 updates and 84.25 at 150 updates. Table 8 concerns an early 1.3B exploratory experiment and was mislabeled as 14B. The final 14B result is:

ModelTotal ↑Quality ↑Semantic ↑
Elastic Forcing 14B83.7984.2481.99

We will correct the affected captions, broken figure/equation references, and duplicated appendices. We will also report VBench alongside VLM and human evaluation, with each result clearly tied to its model and settings.

Reviewer_eKLQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Elastic Forcing post-trains few-step autoregressive video generators without the real- and fake-score models required by DMD. Starting from the same ODE-initialized checkpoint and rollout procedure as Self-Forcing, it instead matches the distribution of the generator's own autoregressive rollouts to a reference video collection by minimizing MMD in frozen self-supervised representation spaces (V-JEPA 2, VideoMAE, and optionally DINOv3). Two estimation techniques make this practical for video: a hybrid Nyström–Monte Carlo estimator of the reference kernel mean, which combines a persistent finite-rank summary of the full reference set with sampled references, and an estimation–differentiation decoupling that evaluates the MMD objective over B=256 fresh rollouts but backpropagates through only b=128 replayed ones with Horvitz–Thompson rescaling. The paper reports VBench Total 84.64 vs. 83.80 for Self-Forcing at 1.3B, post-training of a 14B model in 23.2 hours on 8 GPUs, and adaptation to new styles, concepts, and spatial priors by changing the reference collection.

Critical strengths

A clean and well-motivated replacement for score-based post-training, with correct supporting theory. Removing both auxiliary score networks from the final post-training stage is a genuinely useful simplification, and the two estimation components are developed carefully: the bias–variance decomposition of the hybrid reference estimator (Appendix B.3) and the conditional unbiasedness and covariance of the subsampled replay gradient (Appendix B.4) are derived rather than asserted. The paper is also unusually explicit about the boundaries of its own approximations — the oracle nature of the optimal mixture coefficient, the fact that the training loss is not an absolute MMD value, and the caveat that feature-marginal matching does not identify per-prompt conditional distributions (Appendix B.5).

Critical weaknesses
  1. The provenance of Table 1 is not disclosed, and the central superiority claims rest on numbers the authors apparently did not measure. Table 1 nowhere states whether baseline results were re-measured under the Appendix A.2 protocol or taken from prior work. Cross-checking shows the baseline rows — values and FPS alike, in the same layout — are transcribed from Table 1 of Reward Forcing, which itself mixes scores copied from the Self-Forcing paper with its own re-measurements. This matters quantitatively, not just procedurally. The reference point for the headline improvement, Self-Forcing at 83.80, is Reward Forcing's re-measurement; Self-Forcing's own paper reports 84.31 under its protocol, which would shrink the claimed gain from +0.84 to +0.33. The Rolling Forcing row (81.22/84.08/69.78) comes from that paper's Tables 3–4, which evaluate 5s videos with the official, non-rewritten VBench prompts, whereas Elastic Forcing and the rest of the table use the expanded-prompt protocol; its Semantic score of 69.78, over 10 points below every other row, reflects this prompt mismatch rather than model quality, yet it is ranked in the same table. Relatedly, the paper's own numbers drift by amounts comparable to the claimed gains: the three-encoder configuration is 84.25 in Table 2 but 83.81 in Appendix Table 7, and the hybrid estimator's improvement over pure Nyström estimation is 0.05 points (84.64 vs. 84.59, Table 3). With no variance reported anywhere, the statement that Elastic Forcing outperforms all listed autoregressive baselines is not established by the evidence presented; at minimum the key baselines (Self-Forcing, Rolling Forcing) need to be re-evaluated under the authors' stated protocol, or each row's source and protocol explicitly marked.

  2. Several further claims extend beyond the reported evidence. (i) The "beyond the teacher" framing is broader than what is shown: the main reference collection consists of Wan-generated videos and the generator starts from a teacher-distilled ODE initialization, so teacher knowledge enters through both the samples and the initialization; learning from real-video references is demonstrated only qualitatively (Appendix D.2). (ii) The 14B evidence rests on a VLM evaluation whose scale is not reported — Appendix A.2/H do not state how many prompts or videos per model underlie Table 5 (the human column is 20 prompts/40 videos per Appendix J) — with no VBench, no inference-speed numbers, and a single product-checkpoint baseline of undisclosed training provenance. (iii) No training dynamics are reported: there is no quality-vs-steps curve for MMD post-training, so whether quality saturates or degrades with longer training (e.g., overfitting the 8K reference set, or drifting into encoder-insensitive directions, both live concerns given that the training loss is not an absolute MMD value; Appendix B.1) is unknown, and the choice of 100/150/80 update budgets is unexplained. (iv) The batch-size ablation covers only B ∈ {128, 256}, and the reference-side quantities that determine signal quality — kernel type/bandwidth, λe, αe, number of landmarks R, and the Monte Carlo reference batch size — are not fully reported or ablated. Together with minor internal inconsistencies (e.g., Section 4 reports 8×H200 for the 1.3B run while Appendix Table 6 reports 4×H200 and 100/150 updates; the abstract cites 8×H100 for the 14B run), these gaps make the quantitative conclusions difficult to verify as presented. (v) Finally, the paper does not cite the closest precedent for its central objective: SwD (Starodubcev et al., ICLR 2026) already introduced MMD-based distribution matching in a pretrained model's feature space for few-step diffusion distillation, reported that it works well even in isolation without additional trainable score models, and applied it to image and video models including Wan2.1. Elastic Forcing differs from SwD — external frozen VFM encoders rather than the teacher's own features, whole-rollout marginal matching for autoregressive generation, and the Nyström/decoupling estimators — but this prior work needs to be cited and the novelty delimited against it.

Questions for authors
  1. For each baseline in Table 1, was the number re-measured by you under the Appendix A.2 protocol (946 expanded prompts, 5 seeds) using official checkpoints, or copied from prior work? In particular, can you re-evaluate Self-Forcing (the same ODE-initialized checkpoint you start from) and Rolling Forcing under your exact protocol and prompt set? A per-row provenance/protocol column in Table 1 would resolve this.

  2. Can you reconcile the three-encoder VBench result (84.25 in Table 2 vs. 83.81 in Appendix Table 7), and report run-to-run variability (e.g., standard deviation over training seeds) for the 1.3B main result and the estimator ablation, where the claimed differences are 0.05–0.84 points?

  3. How many prompts and videos per model underlie the VLM scores in Table 5, how many GPT-5.5 calls does the full evaluation require, and are the VLM and human evaluations on the same prompt set? Could you also report the 14B model's VBench (a 14B VBench number, 84.05, already appears in the Appendix C.3 ablation) and its inference throughput?

  4. Could you provide quality as a function of post-training steps (e.g., VBench/VLM at 25/50/100/150/300 updates) and the corresponding objective values? Does MMD training overfit or degrade beyond the reported budgets, and how were those budgets chosen?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eKLQ, Part 1

Thank you for the detailed reproducibility checks and for recognizing the soundness of our estimator and selective-gradient analysis. We now reply to comments and questions.

[W1, Q1] Baseline provenance and a comparison under the same evaluation protocol.

We agree that the provenance of the baseline results should have been explicit. We reused results reported in Reward Forcing and apologize for overlooking the different protocol used for Rolling Forcing. We have now re-evaluated Self-Forcing and Rolling Forcing under the same protocol: 946 expanded VBench prompts and five matched sampling seeds per prompt, giving 4,730 videos per model.

ModelTotal ↑Quality ↑Semantic ↑
Self-Forcing83.9584.7380.83
Rolling Forcing83.0883.7780.30
Elastic Forcing84.6485.4381.48

The improvement persists in Total, Quality, and Semantic under the common protocol. We will replace the corresponding rows and label the source and protocol, including inference-speed measurements.

[W1, Q2] The mixed training steps in Table 2 and the uncertainty of the reported gains.

We clarify that Table 2 mistakenly copied endpoint results from Table 1, mixing different training budgets. As specified in Appendix A, Table 6, the two-encoder model was trained for 100 updates and the final three-encoder model for 150 updates. Table 7 uses a common 100-update checkpoint for the encoder comparison. The VBench total of the three-encoder variant reaches 83.81 on 100 steps and 84.25 on 150 steps. We apologize for the mistake and will correct Table 2 accordingly.

All reported evaluations average five samples per prompt. Across different inference seeds, the standard error of the VBench score for the same model is 0.0935, far smaller than the gap between ours (84.64) and Self-Forcing (83.95).

For the estimator ablation, the hybrid reduces finite-rank reference bias and achieves the best Total, Semantic, and VLM scores in Table 3. Pure Nyström is already strong within our video pipeline. We therefore interpret the small Hybrid–Nyström gap as a component result, while the broader contribution lies in the representation design and the estimation and memory techniques that make MMD practical for video.

[W2(i)] The role of offline teacher data and the meaning of “beyond the teacher.”

We acknowledge that “no teacher” was too broad. Elastic Forcing eliminates online real/fake score models during post-training, removing their recurring computational and memory costs. Although ODE initialization and offline Wan-generated references provide pretrained knowledge, the reference-based objective allows learning distributions beyond the frozen teacher, as demonstrated by our character, style, and spatial-prior experiments.

We further evaluate OpenVid and UltraVideo references, demonstrating that synthetic teacher-generated samples are not required to outperform Self-Forcing. GPT-6 serves as the VLM judge.

Reference sourceVisual Quality ↑Subject Consistency ↑Semantic Consistency ↑Total ↑
Self-forcing4.154.414.494.31
Elastic Forcing (Wan)4.304.594.394.40
Elastic Forcing (OpenVid)4.134.484.114.24
Elastic Forcing (UltraVid)4.344.694.214.39

We use teacher-generated samples in the main experiment to reduce the domain gap from initialization, as in other distributional distillation approaches [LADD, DMAD]. The initialization stage is separate from the reference-based objective and could also be combined with alternative initialization strategies, such as Teacher Forcing intialization [SGF].


[LADD] Sauer et.al., Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation

[DMAD] Yu et.al., DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation

[SGF] Zhuang et.al., Self Gradient Forcing: Native Long Video Extrapolation

AuthorsOpenReview ↗

Response to Reviewer eKLQ, Part 2

[W2(ii), Q3] VLM evaluation size, the mislabeled model in Table 8, and 14B quantitative results.

We provide the missing evaluation details and additional measurements below. We randomly sampled 100 prompts from the VBench evaluation set for VLM evaluation and will release this subset for reproducibility. The human-evaluation set is a subset of these prompts.

Evaluation detailValue
VLM prompts100
Videos per model / total GPT-5.5 calls800
Time to generate a five-second videoApproximately 16 seconds on one H200

Table 8 was mislabeled: it reports an exploratory experiment with an early 1.3B model on a small dataset, used to compare joint text–video and video-only matching. We apologize for the incorrect model label and will correct it. We have also re-measured VBench for the final 14B model:

ModelTotal ↑Quality ↑Semantic ↑
Elastic Forcing 14B83.7984.2481.99

These scores are lower than those of our 1.3B model, which motivated using VLM and human evaluation to assess perceptual improvements at the larger scale. Our 14B model follows the same reference-matching framework, and the Krea comparison follows its released inference pipeline with the same sampling budget. We will include VBench alongside the VLM and human results in the revised Table 5 to provide a more complete assessment.

[W2(iii), Q4] Quality over training, the chosen update budgets, and later-checkpoint behavior.

We agree that training dynamics would make the selected budgets easier to assess. The available three-encoder measurements show improvement from 100 to 150 updates. We selected the budgets by considering generation-quality convergence and computational cost, and will make that criterion explicit alongside the checkpoint results.

Training stepTotalQualitySemantic
069.8174.0652.82
2583.0483.5581.00
5083.3683.9181.17
7583.5284.1381.08
10083.8184.4481.31
15084.2585.0680.99
20083.9084.7880.37
25084.1084.9780.61
30084.1685.0580.59

We will also distinguish the raw per-encoder objective from the normalized training loss. A reference-only term has no direct generator derivative, but the implementation retains it because its value affects the detached normalization denominator. The logged loss should therefore be interpreted as the implemented training surrogate rather than an absolute MMD value.

AuthorsOpenReview ↗

Response to Reviewer eKLQ, Part 3

[W2(iv)] The objective's hyperparameters and the scope of the ablations.

Our goal is to demonstrate strong and consistent performance under a simple, broadly applicable configuration. Therefore, we did not deem an extensive grid search or individual hyperparameter ablations necessary. Instead, the hyperparameters were selected through preliminary small-scale experiments and applied unchanged to the main runs. We believe that achieving strong performance without extensive tuning provides even stronger evidence of our method's effectiveness and generality. We provide the settings below and will include them together in the appendix.

HyperparameterReleased setting
KernelGaussian RBF: ke(X,Y)=exp⁡[−∥X−Y∥F2/(2σe2)]k_e(X,Y)=\exp[-\lVert X-Y\rVert_F^2/(2\sigma_e^2)]
Bandwidth, σe\sigma_eComputed once per encoder/reference bank; fixed during training
Encoder weights / divisorDINOv3:V-JEPA 2:VideoMAEv2 = 10:1:1
Hybrid mixture weightsNyström:MC = 2:1; αe=1/3\alpha_e=1/3
Landmarks, ReR_e1024
Reference minibatch, BrefB_{\mathrm{ref}}128 globally

The bandwidth is the square root of the median positive pairwise squared Frobenius distance over at most 2,048 deterministically selected reference videos. Encoder weights balance the numerical scales of the normalized losses. The recorded 14B settings are:

EncoderFeature matrix per videoEncoder weightRecorded 14B bandwidth σe\sigma_e
DINOv31,024 × 1,02410341.28
V-JEPA 21,024 × 1,02413025.31
VideoMAEv21,280 × 1,02413204.12

These bandwidth values belong to the 14B reference bank; the same construction rule is applied separately to each bank. Table 4 examines the central compute trade-off: increasing the estimation batch improves Total and Quality while keeping the backward batch fixed. The encoder and estimator ablations examine the other principal design choices. Training uses four H200s for 1.3B and eight H200s for 14B; we will correct the inconsistent descriptions.

[W2(v)] SwD's MMD objective and the specific contribution of Elastic Forcing.

We acknowledge SwD as relevant prior work and will discuss it explicitly. SwD matches per-example spatial-token statistics in a diffusion teacher's intermediate features, using linear-kernel MMD in its final formulation. It also reports MMD-only training for Wan2.1, so we do not claim the first standalone use of MMD for video distillation.

Elastic Forcing instead matches populations of complete causal rollouts in external frozen representation spaces, with a persistent reference summary, Monte Carlo correction, and selective differentiation. Our contribution is making standalone reference-defined distribution learning practical for few-step autoregressive video generation, including 14B scaling. We will acknowledge SwD's video and standalone-MMD results while making these distinctions clear.

Reviewer_84D8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a framework for post-training few-step autoregressive video generators using feature-space MMD against reference videos, eliminating online teacher and fake-score networks. Its main contributions are hybrid Nyström–Monte Carlo reference estimation and selective gradient replay to reduce training costs. Experiments include baseline comparisons, component ablations, and VLM/human evaluations. The method improves 1.3B VBench from 83.80 to 84.64 at unchanged inference speed, supports 14B post-training on 8 H200 GPUs, and demonstrates reference-driven style and concept adaptation.

Critical strengths

1, Technical novelty

It addresses a real and significant training bottleneck. Shifting from online maintenance of multiple diffusion networks after autoregressive video training to utilizing fixed reference videos and frozen encoders holds clear resource value. Especially in the 14B experiment, the work goes beyond mere loss function replacement on small models.

2, Experimental validation

Using the same 1.3B architecture and initialization as Self Forcing, the method improves VBench Total from 83.80 to 84.64 while maintaining 17 FPS. Encoder, estimator, and batch-size ablations provide supporting evidence, while the reported 14B experiments extend validation to a substantially larger model and include human evaluation.

Critical weaknesses

1, Technical limitations

The default loss matches video feature marginals. It does not explicitly match text–video correspondence (§3.1; Appendix B.5). So better resemblance to reference videos does not show that new concepts are correctly tied to their prompts. The adaptation experiments lack unseen concept–action combinations, negative prompts, and general-domain retention measurements. This limits evidence for "learning beyond the teacher." The adapted model also gets extra target-domain data the baseline does not.

2, Experimental gaps

Figure 6 does not fully specify matched optimization settings or separate reference-data preparation costs from post-training costs; an OOM result for one Self Forcing implementation does not establish a general resource advantage. The 14B sampling budget is also inconsistent: Figure 1 states four steps, whereas Table 6 specifies five. Human evaluation covers only 20 prompts with one seed each, without confidence intervals for the model difference. These gaps weaken two central claims: better quality and greater computational efficiency.

3, Clarity

a, The abstract reports H100 hardware, whereas §5.1 reports H200

b, Tables 2 and 7 give different results for the three-encoder configuration

c, The main text reports 15 human evaluators while the appendices report 42.

4, Missing related works

Although RDM is cited, it already combines frozen-feature MMD, Nyström reference summaries, fresh generated batches, gradient caching, and multiple encoders. The manuscript does not clearly show which video-specific advances exceed this recipe. It also lacks a fully controlled comparison against straightforward video adaptation. This matters because the strongest remaining distinction, the hybrid estimator, shows only marginal gains. Concurrent ViRDM also warrants explicit discussion.

Questions for authors

1, What specifically distinguishes the proposed training procedure from adapting RDM’s Nyström matching and gradient caching to video? Does the pure-Nyström ablation constitute such a baseline, or are there additional implementation differences?

2, Can you reconcile the conflicting three-encoder results in Tables 2 and 7, H100 versus H200 hardware descriptions, and 15 versus 42 human evaluators?

3, Were Elastic Forcing and Krea evaluated with identical denoising budgets, given the four-step description in Figure 1 and five-step setting in Table 6? For Figure 6, which preparation costs are included, and were the compared methods given comparable memory optimizations and training budgets?

4, How are Table 5’s about 4-point VLM scores derived from the appendix’s 0–100, four-dimension rubric?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 84D8, Part 1

Thank you for recognizing the practical value of removing online score models and our validation at 14B scale. We now reply to comments and questions.

[W1] Whether marginal matching preserves prompt–video correspondence.

We agree that the default objective matches the marginal video distribution. Elastic Forcing also supports joint text–video matching, as implemented and evaluated in Appendix C.3. We adopt marginal matching because joint matching provided no additional benefit in our exploratory experiment, while the default setup maintains strong instruction following in the general-domain evaluation. “Video-only” refers to the matching loss; the generator remains text-conditioned throughout post-training. We will add this discussion to appendix.

[W1] General-domain retention after adaptation and the role of target-domain data.

We provide an additional general-domain VBench evaluation of the Nailoong-adapted model to address retention:

Model, matched re-evaluationTotal ↑Quality ↑Semantic ↑
Self-Forcing83.9584.7380.83
Elastic Forcing after Nailoong adaptation83.6884.7479.43

The adapted model retains substantial general-domain capability while acquiring the target character. Its Total score remains close to Self-Forcing under the same evaluation protocol, with a lower Semantic score. Figure 7 also uses unseen prompts, providing evidence beyond reproducing the training clips.

The key capability demonstrated here is learning directly from target-domain references. Unlike standard DMD with a frozen teacher, which remains tied to the teacher's distribution, Elastic Forcing can learn a new target distribution without adapting the diffusion teacher. The elasticity experiments demonstrate reference-driven adaptation rather than an advantage under equal target-data access. We will incorporate the retention results and clarify this distinction in the revision.

[W4, Q1] Video-specific advances over RDM, the Nyström baseline, and the relation to ViRDM.

We acknowledge that our method builds on RDM's feature-space MMD, Nyström summaries, and gradient caching. Our contribution is to make this approach effective under the representation and computational constraints of causal video, through three complementary components:

  1. Video representations. Our encoder studies motivate dense self-supervised features that preserve both spatial detail and temporal evolution. The tested sparse-feature configurations did not provide an effective video-training signal, and adding image-level constraints alone did not resolve the temporal artifacts in Appendix C.1. Selecting and combining suitable representation spaces is therefore part of the video training procedure.
  2. Comparing more videos than we differentiate. Video rollouts are substantially more expensive than images. Our estimation–differentiation decoupling compares 256 fresh rollouts but backpropagates through only 128, with a gradient scale of 2. This reduces backward computation as well as activation memory. Table 4 isolates the benefit of enlarging the estimation batch while keeping the backward batch fixed.
  3. Hybrid reference estimation. Nyström summarizes the full reference collection but introduces finite-rank approximation bias. The Monte Carlo component reduces this bias, while the persistent summary stabilizes the sampled signal. Table 3 and Figure 12 evaluate the combination quantitatively and qualitatively.
Generated-population settingSamples used for estimationSamples differentiated
RDM's reported FLUX post-training10,24010,240
Elastic Forcing default256128

The pure-Nyström ablation retains our video representations and selective-differentiation pipeline. It isolates the reference estimator rather than reproducing vanilla RDM; we will make this implementation distinction explicit.

We also appreciate the suggestion to discuss concurrent ViRDM, which is already cited. ViRDM emphasizes stochastic exits, staged vector–Jacobian products, and flow-based dynamics regularization. Our method focuses on hybrid reference estimation and selective differentiation, together with multi-encoder supervision, 14B scaling, and reference-driven adaptation. We will add this comparison to the revised discussion.

AuthorsOpenReview ↗

Response to Reviewer 84D8, Part 2

[W2, Q3] Matching the 14B sampling budget and accounting for preparation and training costs.

We clarify that both Elastic Forcing and Krea use Krea's five-step setting. The four-step label in Figure 1 is a typo, for which we truly apologize. Timing begins from the same initialization and records the subsequent distribution-matching stage.

We report reference preparation separately from post-training, for the following reasons.

Reference preparation is a reusable offline cost, unlike recurring teacher evaluations during DMD training. Moreover, Elastic Forcing can leverage publicly available video datasets, such as OpenVid, without generating new references. We use Wan-generated references in the main experiments to reduce the distribution gap from ODE initialization and demonstrate training with real-video references using OpenVid in Appendix D.

We agree that the preparation cost should nevertheless be disclosed separately. The requested details are:

Reference preparationDetails
GeneratorWan2.1-T2V-1.3B
Sampling steps50
Generation hardware and GPU-hoursH200 160 GPU hour
Reference feature / Nyström preparation cost2 GPU hour/0.2 GPU hour

The measured 1.3B comparison and 14B post-training in 23.2 hours on eight H200s support the resource benefit. The OOM marker records the tested Self-Forcing implementation under the most aggressive offloading strategies. Krea reports that naive Self Forcing implementation causes OOM on 64 H100 GPUs; we will specify those configurations in the revision. For figure 6, we use the exact same memory optimization strategies and training configs (flash attention 2, same batch size), and different offloading strategies balance the speed-memory tradeoff.

[W2] Human-evaluation coverage and uncertainty in the model difference.

We provide the study size and the paired model difference below. The human study involved 42 annotators evaluating 20 matched prompt pairs and favored Elastic Forcing over Krea (Table 5).

Additional evaluation informationResult
Paired mean difference, Elastic Forcing − Krea+0.300
Prompts in the broader VLM evaluation100

The human study provides complementary evidence independent of the training encoders and the VLM judge. We will compute the confidence interval over prompt pairs, rather than treating repeated ratings of the same videos as independent pairs.

[W3, Q2] The cause of the table discrepancy and the correct hardware and rater counts.

We apologize for a reporting mistake in Table 2: it copied the endpoint results from Table 1 and therefore mixed 100- and 150-update checkpoints. Table 1 reports the two-encoder model at 100 updates and the final three-encoder model at 150 updates. Table 7 compares the variants at a common 100-update checkpoint. The three-encoder scores are consequently 83.81 at 100 updates and 84.25 at 150 updates. We will correct Table 2 to use the common-update comparison and label the final-model checkpoints separately.

Training uses four H200s for 1.3B and eight H200s for 14B. The human evaluation involved 42 annotators. The split-half rater correlation is 0.689; we will clarify how the split was constructed in the evaluation details. We will correct the inconsistent hardware and participant counts throughout the revision.

[Q4] Converting the appendix's 0–100 VLM rubric to the scores in Table 5.

We clarify that Table 5 rescales the original 0–100 ratings to a five-point display scale, while retaining the original scores for aggregation. The conversion is a linear mapping, s=1+4r/100, where r denotes the original rating and s the displayed score. Appendix H specifies the equal-weight mean of four dimensions. We will state the exact conversion formula and include the temporal-coherence column omitted from Table 5. The ablation VLM scores use a separate relative-ranking protocol.

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

Related papers

Discussion (0)

Sign in to comment.