acceptodds
Under review as a conference paper at ICLR 2027

RepForcing: Representation-Space Initialization for Autoregressive Video Distillation

Abstract

Distilling pretrained video diffusion models into few-step autoregressive (AR) generators typically combines student initialization with distribution matching distillation (DMD). An intermediate AR teacher can provide causally compatible targets for ODE initialization. However, causally compatible targets alone do not ensure high-quality generation under a short sampling schedule. We analyze the posterior-mean predictions favored by direct latent regression and empirically find that causal ODE initialization alleviates blur but still leaves fine detail unresolved. To address this limitation, we propose Representation Forcing, which learns few-step causal initialization through representation-space supervision rather than intermediate-teacher trajectory regression. The student is trained with teacher forcing by comparing clean predictions and training targets through normalized features of a frozen video diffusion transformer. This phase reuses the synthetic videos generated by a bidirectional model for AR teacher training and replaces both teacher training and causal ODE distillation, without generating additional causal ODE trajectories. The subsequent asymmetric DMD procedure remains unchanged. Empirical results show that, at four sampling steps, RepForcing better preserves generation quality than intermediate-teacher-based initialization before DMD. After DMD, our four-step chunk-wise model improves VBench Total from 84.04 to 84.66 (+0.62 points) and VisionReward from 6.326 to 10.074 (+59.25%) over the reported causal-ODE-based baseline.

Primary area:
generative models
Keywords:
autoregressive video generation, perceptual supervision, diffusion distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_oULGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors target the initialization stage of few-step autoregressive (AR) video distillation, where Causal Forcing trains an intermediate AR teacher and distills its causal ODE trajectories before asymmetric DMD. The authors argue that latent regression favours posterior-mean predictions, and observe that ODE initialization still leaves fine detail unresolved at four steps. RepForcing replaces the teacher and the ODE stage with a single teacher-forced phase: the causal student (Wan2.1-1.3B) predicts clean blocks, and the loss is a token-wise cosine distance between features of the predicted and target latents, read from block 8 of a frozen Wan2.1-14B. DMD is unchanged and no ODE trajectory dataset is generated. After DMD, the four-step model reports the best VBench Total, Quality, VisionReward and Instruction Following among distilled AR models in both settings.

Critical strengths
  • This work presents a genuine simplification: it deletes the intermediate teacher and the ODE trajectory dataset while leaving the student, the data and DMD untouched.
  • The final numbers are strong in both settings, and the reward gains are large rather than marginal (VisionReward 6.326 to 10.074 chunk-wise, 6.798 to 10.598 frame-wise). The comparison reuses Causal Forcing's dataset, backbone and evaluation prompts, so it is better controlled than a typical cross-paper table.
  • The paper reports where it loses and does not overclaim: it says plainly that deterministic ODE targets alone do not account for the few-step quality gap.
Critical weaknesses
  • The matched harness behind Table 2 is never pointed at the two comparisons the thesis requires: a latent-MSE Stage 1, and an ODE or CD initialization.
  • Pipeline simplification is the headline benefit, yet no Stage-1 hyperparameters or training costs are reported.
  • Novelty is modest, essentially a perceptual loss with a frozen video DiT as the feature network. The related work is honest about this, and on its own I would not weigh it against the paper.
Questions for authors
  1. Can you report pre-DMD metrics (VBench Quality, Semantic, Dynamic Degree, VisionReward at 4, 8 and 12 steps) for RepForcing and at least one intermediate-teacher initialization?
  2. What is your DMD configuration (teacher, guidance scale, update budget), and does it match the runs behind the Causal Forcing and Causal Forcing++ numbers you quote?
  3. What are the Stage-1 hyperparameters and total training cost, and how do they compare with teacher training plus ODE generation? Are target features cached?
  4. Wan2.1-14B is the extractor, the DMD teacher and, via the bidirectional Wan model, the source of the training videos. Does the benefit survive when the extractor is decoupled from the teacher and the data generator?
6 comments
AuthorsOpenReview ↗

Our Response to Reviewer oULG (1/2) [W1, Q1, W2, Q2, Q3]

We thank the reviewer for the detailed summary and for recognizing that the comparison reuses Causal Forcing's data, backbone and prompts, and that the paper does not overclaim.

W1 / Q1. The matched harness is not pointed at a latent-MSE Stage 1 or at an ODE/CD initialization; report pre-DMD metrics at 4, 8 and 12 steps. We provide quantitative pre-DMD comparisons between RepForcing and the causal-ODE and CD initializations on the full evaluation set at 4, 8 and 12 sampling steps. RepForcing consistently improves VBench Total, Quality and VisionReward over both baselines across these sampling schedules, supporting its generation-quality and reward gains before DMD. The detailed results are presented in Table R3 (reply to Reviewer NQeq, W1 / Q1).

W2 / Q2 / Q3. Training configuration, cost, and feature caching.

DMD configuration (Q2). Yes, our DMD stage matches the runs behind the quoted baselines: RepForcing follows the asymmetric DMD recipe of Self Forcing, Causal Forcing and Causal Forcing++, sharing the teacher, guidance scale, auxiliary score model, optimizer, schedule, batch and training prompts, and only the initialization differs. The post-DMD differences in Table 1 therefore reflect the initialization rather than the distillation recipe. The table below specifies our four-step configuration and update budget.

SettingSelf Forcing / Causal Forcing (released configs)RepForcing
Teacher (real score)Wan2.1-T2V-14B, guidance 3.0same
Auxiliary (fake) scoreWan2.1-T2V-1.3Bsame
Generator / critic learning rate2×10⁻⁶ / 4×10⁻⁷, β = (0, 0.999)same
Critic updates per generator update5same
Four-step schedule[1000, 750, 500, 250], warped, shift 5.0same
EMA0.99 from step 200same
Global batch64same
Training promptsVidProM, filtered and extended (released with Self Forcing)same
Training lengthnot fixed in the release; ≤ 1K iterations recommended at batch 64750 iterations

Stage-1 hyperparameters and cost. We use AdamW with learning rate 2×10⁻⁶, β = (0, 0.999), weight decay 0.01, global batch 64 videos (16 GPUs × 1 video per GPU × 4 accumulation steps), gradient-norm clipping at 10, bf16 with FSDP and gradient checkpointing, no EMA, and 3,000 iterations. Each block draws an independent noise level from the 1000 training timesteps with timestep shift 5.0. The extractor is the frozen Wan2.1-14B with native bidirectional attention, read at block 8 (one-based) with τ = 0 on clean latents; the loss is the token-wise cosine distance. RepForcing's pre-DMD training costs 660 A800 GPU-hours, compared with 905 GPU-hours for Causal Forcing's AR-teacher training and additional ODE stage, reducing the total cost by 27.1%. Peak memory per GPU is 54 GB for RepForcing and 46 GB for Causal Forcing. The detailed comparison is presented in Table R6 (reply to Reviewer NQeq, W3 / Q3); DMD is not included.

Target feature caching. Target features are not cached in our implementation; they are recomputed at every step under stop-gradient. They can be cached, since the target branch is detached and its input (the clean latent at τ = 0) is fixed.

We will include these training settings and cost details in the revision.

AuthorsOpenReview ↗

Our Response to Reviewer oULG (2/2) [W3, Q4]

W3. Novelty. We agree that the loss is a perceptual-style objective with a frozen video DiT as the feature network, and we appreciate the reviewer's note that this alone should not count against the paper. The contribution is showing that, for few-step causal autoregressive video initialization, changing the regression metric is enough to replace the intermediate AR teacher and the ODE trajectory stage, together with the analysis and design findings that make it work: the condition under which the representation loss rejects averaged predictions (new Corollary, Appendix C.3), video-diffusion features outperforming DINOv2 and InternVid, the role of intermediate depth (Table 2), and the irrelevance of the extractor's attention mask (Table R5).

Q4. Does the benefit survive when the extractor is decoupled from the teacher and the data generator? We first clarify that the shared training videos are generated by bidirectional Wan2.1-1.3B, whereas the feature extractor and DMD teacher are Wan2.1-14B.

Table 2 already decouples the extractor from the DMD teacher: with the Wan2.1-14B teacher fixed, DINOv2 and InternVid extractors perform worse, so part of the benefit comes from using video-diffusion features rather than from sharing a model with the teacher. Table R5 adds that the benefit does not depend on how the extractor attends over time. Replacing the extractor with a video DiT from a different family requires re-encoding through that model's VAE and is beyond what we could run in the rebuttal period; we list it, together with training on videos from another generator, as a limitation and future work in the revision.

Summary. We thank the reviewer again for the careful reading and for recognizing the simplification and the honesty of the paper. Following the reviewer's suggestions, the matched harness now covers the latent-MSE AR-diffusion model and the causal-ODE and CD initializations before DMD (Tables R1 and R3), the DMD configuration is shown to match the quoted baselines, Stage-1 hyperparameters, cost and caching are fully specified (Table R6), and the extractor-coupling question is answered with Table 2 and a new attention-mask ablation (Table R5). We believe these additions resolve the reviewer's main concerns (W1 and W2), and we would be grateful if the reviewer would consider updating the score accordingly. We remain available for any further questions.

Dear Reviewer oULG,

We hope this message finds you well. We understand that this is a busy period, and we sincerely appreciate the time you have devoted to reviewing our submission.

We have posted our responses to your comments and uploaded a revised manuscript that incorporates your suggestions. As the discussion period will close on November 18, we would be grateful if you could take a moment to let us know whether our responses have addressed your concerns. If anything remains unclear, we would be glad to provide further clarification or additional results.

Thank you again for your thoughtful feedback.

Best regards,

RepForcing Authors

Reviewer_NQeqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RepForcing, which replaces intermediate AR-teacher training and causal ODE distillation with a teacher-forced initialization objective in the feature space of a frozen video DiT. The causal student is subsequently refined using asymmetric DMD. The authors report that RepForcing preserves more visual detail with four sampling steps and generally outperforms recent autoregressive video distillation methods in both chunk-wise and frame-wise generation.

Critical strengths
  • This paper is well written and easy to follow.
  • The simpler initialization pipeline is practically appealing.
  • The method shows promising visual detail and VisionReward gains.
Critical weaknesses
  1. The central claim of better pre-DMD initialization relies mainly on the qualitative example in Fig2, a quantitative comparison across prompts is missing.
  2. The analysis does not establish that feature regression avoids conditional averaging. Moreover, the bidirectional feature extractor uses future frames when supervising a causal student.
  3. Total training cost is not reported, making the claimed simplification difficult to assess computationally.
Questions for authors
  • Can the authors report quantitative pre-DMD results over the full evaluation set, ideally alongside a matched causal-ODE baseline?
  • How much does the loss or gradient for an early block change when future frames are perturbed? Does bidirectional feature supervision introduce a measurable causal mismatch?
  • What are the total training GPU-hours and peak memory compared with Causal Forcing?
6 comments
AuthorsOpenReview ↗

Our Response to Reviewer NQeq (1/3) [W1, Q1]

We thank the reviewer for the clear and actionable questions.

W1 / Q1. Quantitative pre-DMD comparison over the full evaluation set with a matched causal-ODE baseline. We supplement the qualitative comparison with quantitative pre-DMD evaluations on the full evaluation set, comparing RepForcing with causal-ODE and CD initializations at 4, 8 and 12 sampling steps using the same sampler. RepForcing improves VBench Total, Quality and VisionReward over both baselines at all three step counts. At four steps, it improves Total by 2.21 points, Quality by 2.03 points and VisionReward by 23.3% over causal ODE; compared with CD, the gains are 1.11 points, 1.37 points and 19.0%, respectively. The VisionReward improvement over causal ODE remains 19.8% at eight steps and reaches 28.3% at twelve steps. These results quantitatively support the generation-quality and reward gains before DMD. The full results are presented in Table R3.

Table R3. Pre-DMD generation, chunk-wise, same sampler for all models. Each cell lists VBench Total / Quality / Semantic / Dynamic Degree / VisionReward. Higher is better; bold marks the best initialization for each metric and step count.

Initialization4 steps8 steps12 steps
Causal Forcing causal-ODE init (released)80.13 / 81.14 / 76.10 / 77 / 3.84780.24 / 80.84 / 78.83 / 83 / 4.28480.54 / 81.62 / 76.21 / 76 / 4.311
Causal Forcing CD init (released)81.23 / 81.80 / 78.92 / 71 / 3.98981.37 / 82.01 / 78.78 / 78 / 4.30782.54 / 83.62 / 78.21 / 80 / 4.426
RepForcing (3,000 updates)82.34 / 83.17 / 79.00 / 79 / 4.74582.72 / 83.97 / 77.73 / 81 / 5.13282.91 / 84.24 / 77.58 / 74 / 5.532
AuthorsOpenReview ↗

Our Response to Reviewer NQeq (2/3) [W2]

W2 (part 1). The analysis does not establish that feature regression avoids conditional averaging. The representation-regression optimum is the realizable prediction whose features are closest to the conditional feature mean, which need not coincide with the coordinate-wise latent mean. However, this result alone does not establish that representation regression always avoids averaged predictions. We consider a posterior with two equally probable modes to specify when the representation loss prefers a mode over the averaged prediction.

Corollary (two-mode posterior). Let the posterior of a block place equal mass on two realizable continuations aa and bb, let m=(a+b)/2m = (a+b)/2 be their latent average, and write d(⋅,⋅)d(\cdot,\cdot) for the distance between unit-normalized features. The representation loss is 12(d(m,a)2+d(m,b)2)\frac{1}{2}(d(m,a)^2 + d(m,b)^2) at the average and 12d(a,b)2\frac{1}{2}d(a,b)^2 at either mode, so a mode attains a strictly lower loss than the average if and only if d(a,b)2<d(m,a)2+d(m,b)2d(a,b)^2 < d(m,a)^2 + d(m,b)^2; in the symmetric case, iff d(m,a)>d(a,b)/2d(m,a) > d(a,b)/\sqrt{2}. For latent ℓ2\ell_2 (identity features) d(m,a)=d(a,b)/2d(m,a) = d(a,b)/2, so the average always wins, which recovers Proposition 1; a representation that places the blurred average farther than d(a,b)/2d(a,b)/\sqrt{2} from each mode prefers a sharp mode instead.

This turns "need not coincide with the mean" into a checkable condition on the feature map, and it is what our design choices act on. Sec. 4.2 already argues that shallow features are close to an isometry of the latent space, where the average lies at half the mode separation and therefore wins, which matches the drop at relative depth 1/8 in Table 2; intermediate-depth features of a video diffusion model, which is trained to tell clean videos from corrupted ones, beat DINOv2 and InternVid. We do not claim that averaging is avoided in general, only that it is penalized whenever the representation separates averages from modes; empirically, Table R1 shows markedly smaller single-step errors than the latent-regression-based initializations at high noise, where the posterior is most spread out.

We will include this corollary and its proof in the revision (new Appendix C.3).

W2 (part 2). The bidirectional extractor uses future frames when supervising a causal student.

The extractor plays the role of a frozen perceptual metric, like LPIPS in image models: it receives the student's clean prediction and the clean target, never feeds anything into the student, and is discarded before DMD and inference. The student remains strictly block-causal at training and test time. The relevant question is therefore whether the metric's attention pattern changes the supervision the student receives. We measured this directly and found that it changes it only marginally, at the level of the loss, of the gradient, and of the trained model.

Sensitivity of the loss (Table R4). For early blocks i, we replace only the future blocks (> i) and measure the change of block i's supervision. The prediction is the teacher-forced clean prediction of the released Causal Forcing AR-diffusion model, so it has realistic errors.

Table R4. Effect of future blocks on block-i supervision (bidirectional extractor, block 8, 32 clips).

Perturbation of blocks > iEffect on block i
Target: future replaced by another clip'sloss shift 0.012 (i=1) decreasing to 0.003 (i=6), i.e. 1–4% of the distance between two unrelated clips (0.30–0.39)
Prediction: future replaced by ground truthgradient cosine 0.92–0.97
Prediction: future replaced by another clip'sgradient cosine 0.82–0.92
Prediction: future replaced by large Gaussian noise (σ = 0.5 std)gradient cosine 0.67–0.81

Even with an unrelated future, the block-i gradient keeps its direction, and the target moves by only a few percent of the scale that separates different videos.

Agreement of the training signal. At the same predictions, the full training-loss gradient of the bidirectional extractor and that of a block-causal extractor (the student's 3-frame block mask) have a cosine similarity of 0.85–0.89 on blocks 2–7, whereas either has a cosine of only 0.34–0.42 with the latent-ℓ2 gradient. The attention mask of the metric thus changes the supervision far less than the choice of metric itself. (On the first block all variants agree less, 0.21–0.28, including two causal variants with each other, so this is a property of the first block, not of bidirectional attention.)

AuthorsOpenReview ↗

Our Response to Reviewer NQeq (3/3) [Q2, W3, Q3]

Q2. Trained block-causal extractor ablation.

Trained ablation (Table R5). We trained the block-causal variant with every other setting unchanged (block 8, τ = 0, data, seed, global batch 64, schedule) and evaluated both initializations on the 200 held-out clips of Table R1.

Table R5. Bidirectional vs. block-causal extractor: single-step (teacher-forced) clean prediction on the same 200 held-out clips of DBiD_{\mathrm{Bi}} as Table R1, before DMD. Both variants are re-trained with our released code on the remaining clips for 2,400 updates with identical settings; the bidirectional rows are the RepForcing (2,400 updates) model of Table R1.

Extractor attentiont=1000t=750t=500t=250
Latent MSEbidirectional (default)0.5320.2220.1090.0466
block-causal0.5350.2210.1080.0462
Decoded PSNR / LPIPSbidirectional (default)12.33 / 0.56615.69 / 0.37518.55 / 0.23222.08 / 0.119
block-causal12.28 / 0.57015.70 / 0.37818.56 / 0.23222.09 / 0.119

The two variants differ by less than 1% in latent error and by at most 0.05 dB in PSNR, at every noise level, and their training curves coincide at 250, 500, 1,000 and 2,400 updates (Appendix E). Each variant is marginally better only when scored with its own extractor (e.g. block-causal 0.151 vs. bidirectional 0.155 under the block-causal metric at t = 1000; bidirectional 0.154 vs. 0.156 under the bidirectional metric), as expected. The block-causal variant is also about 17% faster per iteration. We conclude that bidirectional attention in the extractor is not what drives the results; it is the video-diffusion representation used as the metric, consistent with the extractor ablation in Table 2. We will state this in the revision and list the block-causal extractor as an equally valid, cheaper option.

W3 / Q3. Total training GPU-hours and peak memory compared with Causal Forcing.

Following Causal Forcing, we reuse its first-stage dataset DBiD_{\mathrm{Bi}}: approximately 3K videos synthesized by bidirectional Wan from VidProM prompts. We compare pre-DMD training costs on A800 GPUs using this shared dataset. RepForcing requires only one representation-supervised AR training stage, reducing the total cost from 905 to 660 GPU-hours, a reduction of 27.1%, while eliminating the additional ODE stage and its trajectory storage. The cost and peak-memory comparison is presented in Table R6.

Table R6. Pre-DMD training cost on A800 for approximately 3K videos. DMD is not included.

StageCausal ForcingRepForcing
AR model training (3,000 updates)470 GPU-h (AR teacher)660 GPU-h (representation-space supervision)
Additional ODE trajectory generation + distillation435 GPU-h—
Total pre-DMD training905 GPU-h660 GPU-h
Peak memory per GPU46 GB54 GB (FSDP, gradient checkpointing)
Extra storage for trajectory data71 GiBnone

For approximately 3K videos and the same 3,000 updates, Causal Forcing's AR-teacher training costs 470 A800 GPU-hours, followed by ODE trajectory generation and distillation costing 435 GPU-hours and requiring 71 GiB of extra storage. Its total pre-DMD training cost is therefore 905 GPU-hours. RepForcing uses a single representation-supervised AR training stage of 3,000 updates, costing 660 A800 GPU-hours, and directly uses the resulting model for DMD without an additional ODE initialization stage. Peak memory is 46 GB per GPU for Causal Forcing and 54 GB for RepForcing.

RepForcing trains on 16 A800 GPUs with a global batch of 64 videos. The extractor keeps only the first 8 of Wan2.1-14B's 40 blocks. No separate intermediate teacher is trained and no trajectory dataset is generated or stored.

Summary. We thank the reviewer again for the precise questions. We now report (i) quantitative pre-DMD results on the full evaluation set, in which RepForcing improves VBench Total, Quality and VisionReward over both the causal-ODE and the CD initializations at 4, 8 and 12 steps (Table R3); (ii) a corollary that makes precise when the representation loss rejects averaged predictions; (iii) direct measurements showing that future frames have only a negligible effect on the supervision of the bidirectional extractor, together with a trained block-causal ablation that matches the default (Tables R4–R5); and (iv) a 27% lower pre-DMD training cost than Causal Forcing (Table R6). As these results address each of the reviewer's points with new evidence, we would sincerely appreciate it if the reviewer would consider raising the score. We are glad to provide any further details.

Reviewer_fpMzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RepForcing, an initialization method for AR video distillation. The authors observe that Causal ODE initialization alleviates blurriness but still struggles to preserve fine details under short sampling schedules. Instead of training an intermediate AR teacher and distilling its trajectories, RepForcing directly trains causal students by applying representation-space supervision. The loss is defined as token-wise cosine similarity between features extracted from the student's predicted clean latents and the target clean latnets. Experiments are conducted on several long video methods using standard evaluation metrics.

Critical strengths

The proposed method is quite simple and straightforward. It eliminates the need for intermediate AR teacher training and additional causal ODE trajectory generation. Moreover, since the feature extractor is used only during initialization, no additional model is required at inference time. This simplicity is a practical advantage of the proposed approach.

The ablation results are interesting. Although the proposed method uses representations extracted from the pretrained teacher, it achieves better performance than aligning with strong foundation models such as DINOv2. This finding is meaniful, as it suggests that internal representations from the video diffusion teacher may provide more effective supervision for initializing an AR video generator.

Critical weaknesses

While I find the analysis in sec4.1 interesting, I am not fully convinced that it sufficiently motivates the proposed representation-space initialization in sec4.2. The analysis shows that latent regression favors posterior-mean predictions, but it is unclear why representation-space supervision is an appropriate solution to this issue. Furthermore, causal ODE initialization uses deterministic targets. So I think it already avoids the conditional ambiguity discussed in the analysis. Therefore, I would appreciate a clearer explanation of how the proposed representation-space objective specifically addresses the limitations of causal ODE initialization under few-step sampling.

Although the paper provides qualitative comparisons through sampled frames, full videos are not available. Since this paper focuses on autoregressive video generation, it is difficult to assess temporal consistency or motion quality. I believe providing video demonstrations is important for a more comprehensive evaluation.

Questions for authors

Q1. Regarding W1, could the authors provide a clearer explanation of why representation-space supervision is more effective than causal ODE initialization? Fig2 partially shows that RepForcing better preserves visual details, particularly under four-step generation. However, since causal ODE initialization already uses deterministic targets, I am curious about what specific limitation of this approach is addressed by representation-space supervision. I would appreciate further theoretical or empirical evidence connecting the analysis in sec4.1 to the proposed method in sec4.2

Q2. While the proposed method shows promising results under four-step generation, which is a standard setting, it remains unclear how well it generalizes to one-step or two-step generation. Could the authors provide results under these sampling schedules? Given that the method is motivated by the limitations of regression-based initialization under few-step sampling, this is an important setting to investigate. In particular, it would be interesting to see whether representation-space supervision continues to outperform causal ODE initialization as the number of sampling steps decreases.

Q3. Could the authors provide training loss curves and convergence analysis for the proposed representation-space initialization? I think one of the main advantages of RepForcing is simplifying the initialization pipeline. So, it would be useful to understand whether this also leads to improved training efficiency.

9 comments
AuthorsOpenReview ↗

Our Response to Reviewer fpMz (1/2) [W1, Q1]

We thank the reviewer for the positive assessment of the method's simplicity and of the representation ablation.

W1 / Q1. Why is representation-space supervision more effective than causal ODE initialization, given that ODE targets are deterministic?

We agree that deterministic ODE targets remove the conditional target ambiguity of Proposition 1; Sec. 4.1 states that target determinism alone does not ensure high-quality few-step generation. The remaining gap therefore comes from what ODE initialization does not change, and we identify two factors.

(a) The target source. ODE endpoints are produced by an intermediate AR teacher trained on the same videos, so detail that this teacher fails to produce is absent from the targets. RepForcing regresses directly to the training videos.

(b) The regression metric under limited capacity. ODE targets are a deterministic function of the student's input, so the target itself is unambiguous. A capacity-limited few-step student, however, cannot resolve every input of the teacher's flow map: inputs it cannot tell apart act like an ambiguous condition, and for them the ℓ2-optimal compromise is again the average of their endpoints (Proposition 1 applied to the student's effective inputs). The representation loss rejects such averages whenever the feature space places them far from the individual endpoints; the new Corollary in Appendix C.3 (reply to Reviewer NQeq, W2) gives the exact condition. This is why the regression metric matters even with deterministic targets, and why the effect is strongest at high noise, where the student has the least information to separate continuations.

Table R1 measures the outcome directly: under identical noise, conditioning and clean prefixes, the single-step clean prediction of the RepForcing initialization is closer to the held-out target than that of the intermediate-teacher models, and the advantage is largest at the high-noise steps where few-step sampling has the least information.

Table R1. Single-step (teacher-forced) clean prediction on 200 held-out clips of DBiD_{\mathrm{Bi}}, before DMD. Lower is better. For this controlled comparison RepForcing is re-trained with our released code on the remaining clips for 2,400 updates; the released Causal Forcing models are used as is. Every block is noised to the same t from the four-step schedule, with noise and prompts fixed per clip.

Initializationt=1000 latent MSEt=750t=500t=250t=1000 rep. losst=750t=500t=250
Causal Forcing AR diffusion0.6170.2450.1100.04450.1870.1210.0730.036
Causal Forcing causal-ODE init0.7150.2690.1250.05160.1940.1160.0710.037
RepForcing (2,400 updates)0.5320.2220.1090.04660.1540.0930.0590.032

"Rep. loss" is the block-8 Wan2.1-14B cosine distance used in training; latent MSE is not optimized by RepForcing and serves as a neutral metric. The released models were trained by their authors on data that may include these clips, which only favours them. The only cell where RepForcing is not best is the latent MSE at the lowest noise level, where the latent-trained AR diffusion model is marginally better (0.0445 vs. 0.0466); there the prediction is nearly determined by the input and latent regression is at its strongest. We have also toned down Sec. 4.1–4.2: Appendix C.2 shows only that the representation optimum need not coincide with the latent mean.

AuthorsOpenReview ↗

Our Response to Reviewer fpMz (2/2) [W2, Q2, Q3]

W2. Full videos. We agree that sampled frames alone are insufficient for assessing temporal consistency and motion quality. We provide video comparisons between RepForcing and other baseline methods, allowing reviewers to examine visual details, temporal consistency and motion throughout the videos rather than only at selected frames. The comparison videos are presented on an anonymous demo website at https://repforcing-demos.pages.dev/.

Q2. One- and two-step generation. We provide additional evaluations of post-DMD one- and two-step generation, comparing RepForcing with the corresponding few-step versions of Causal Forcing++ under the same schedules (including Causal Forcing++'s four-step first chunk). RepForcing retains advantages in generation quality, VisionReward and Instruction Following under both sampling schedules. The results are presented in Table R2.

Table R2. Post-DMD one- and two-step frame-wise generation. Higher is better; bold marks the better model at each step count.

ModelTotal ↑\uparrowQuality ↑\uparrowSemantic ↑\uparrowDynamic. ↑\uparrowVision. ↑\uparrowInstruct. ↑\uparrow
Causal Forcing++ (1-step)83.3584.5078.75665.41238
Causal Forcing++ (2-step)84.1484.8981.13646.66151
RepForcing (1-step)83.6684.7479.36608.71559
RepForcing (2-step)84.4585.2481.306710.40276

RepForcing improves VBench Total and Quality over Causal Forcing++ under both one- and two-step generation: Total increases from 83.35 to 83.66 at one step and from 84.14 to 84.45 at two steps, while Quality increases from 84.50 to 84.74 and from 84.89 to 85.24, respectively. VisionReward increases from 5.412 to 8.715 at one step and from 6.661 to 10.402 at two steps, and Instruction Following increases from 38 to 59 and from 51 to 76. Semantic improves from 78.75 to 79.36 at one step and from 81.13 to 81.30 at two steps; Dynamic Degree is lower at one step (60 vs. 66) and higher at two steps (67 vs. 64). These results show that the final distilled model retains quality and reward gains under shorter sampling schedules.

Q3. Loss curves and convergence. Fig. R2 plots the Stage-1 representation loss of our chunk-wise run over the first 3,000 iterations (per-iteration values and a 100-iteration trailing mean; the per-iteration loss is noisy because every iteration samples new noise levels). The objective converges quickly and smoothly. The windowed loss drops from 0.081 over the first 50 iterations to 0.064 over iterations 51–100 and 0.056 over iterations 101–200, i.e. about half of its total decrease within 100 iterations and three quarters within 200; it reaches about 90% within 500 iterations and 97% within 1,500, after which it only refines slowly (0.048 over iterations 2,501–3,000). Because this single stage replaces both the AR teacher and the ODE stage, and converges early, the pipeline is also cheaper: RepForcing reaches its final initialization with one 3,000-update stage, whereas Causal Forcing needs an AR teacher of the same length plus an ODE stage, so its total pre-DMD cost is 27% lower (Table R6, reply to Reviewer NQeq).

Summary. We thank the reviewer again for the constructive questions. We have (i) added full videos on an anonymous website, (ii) explained, with a new corollary and the controlled single-step comparison in Table R1, why the representation loss helps even with deterministic ODE targets, (iii) reported one- and two-step results in which RepForcing stays ahead of Causal Forcing++ on VBench Total, Quality, VisionReward and Instruction Following, and (iv) provided loss curves showing fast and stable convergence. We hope these additions address the reviewer's concerns about the motivation and the evaluation, and we would be grateful if the reviewer would consider raising the score. We are happy to clarify anything further during the discussion period.

AuthorsOpenReview ↗

Kindly reminder

Dear Reviewer fpMz,

Thank you again for the time and effort dedicated to reviewing our work.

We have submitted a detailed response to your initial comments and have revised the manuscript accordingly. As the discussion period is progressing, we wanted to respectfully follow up to ensure that our response has sufficiently addressed your concerns. Please let us know if there are any remaining questions or if further clarification is needed on any specific points. We remain available to answer any further questions you may have before the discussion period closes.

Best regards,

RepForcing Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Summary of Revisions

We thank all reviewers for their constructive comments. We have revised the manuscript and added new analyses and experiments to address every concern. The major updates are summarized below; tags such as #fpMz-W1 refer to the corresponding reviewer comment.

1. [Main Paper Sec. 4.1–4.2 & Appendix C.3] Why Representation Supervision Helps Beyond Deterministic ODE Targets (#fpMz-W1/Q1, #NQeq-W2)

  • Two sources of the remaining gap: Deterministic ODE targets remove target ambiguity but leave (a) the target source, an intermediate teacher whose missing details are absent from its targets, and (b) the regression metric, under which a capacity-limited few-step student still averages over inputs it cannot distinguish.
  • New Corollary (Appendix C.3): An exact condition under which the representation loss prefers a sharp mode over the averaged prediction, recovering Proposition 1 as the latent-ℓ2\ell_2 case.
  • Controlled single-step comparison (Table R1): On 200 held-out clips, RepForcing yields the lowest error at high noise (latent MSE at t=1000t=1000: 0.532 vs. 0.617 for AR diffusion and 0.715 for causal-ODE init). We also toned down the claims in Sec. 4.1–4.2.

2. [Table R3] Quantitative Pre-DMD Comparison (#NQeq-W1/Q1, #oULG-W1/Q1)

  • Matched comparison at 4/8/12 steps: RepForcing improves VBench Total, Quality and VisionReward over both causal-ODE and CD initializations at all step counts. At four steps, it gains +2.21 Total, +2.03 Quality and +23.3% VisionReward over causal ODE; the VisionReward gain reaches 28.3% at 12 steps.

3. [Tables R4–R5 & Appendix E] Bidirectional Extractor and Causal Mismatch (#NQeq-W2/Q2, #oULG-Q4)

  • Sensitivity to future frames (Table R4): Replacing the future target blocks with another clip shifts early-block supervision by only 1–4% of the distance between unrelated clips, and even with an unrelated future in the prediction, the block gradient keeps its direction (cosine 0.82–0.92).
  • Trained block-causal ablation (Table R5): A block-causal extractor matches the default within <1% latent error and ≤0.05 dB PSNR at every noise level while running 17% faster; we now list it as an equally valid option.

4. [Table R2] One- and Two-Step Generation (#fpMz-Q2)

  • Post-DMD results: Versus Causal Forcing++, RepForcing improves VBench Total (83.35 → 83.66 at one step; 84.14 → 84.45 at two), VisionReward (5.412 → 8.715; 6.661 → 10.402) and Instruction Following (38 → 59; 51 → 76).

5. [Table R6 & Fig. R2] Training Cost, Hyperparameters and Convergence (#NQeq-W3/Q3, #oULG-W2/Q3, #fpMz-Q3)

  • Cost: Pre-DMD training drops from 905 to 660 A800 GPU-hours (−27.1%), without the ODE stage or its 71 GiB trajectory storage (peak memory 54 vs. 46 GB per GPU).
  • Specification and convergence: Full Stage-1 hyperparameters are reported; the loss reaches about half of its total decrease within 100 iterations and 90% within 500.

6. [DMD Settings] DMD Configuration (#oULG-Q2)

  • Identical recipe: RepForcing shares the teacher, guidance, auxiliary score model, optimizer, schedule, batch and prompts with Self Forcing / Causal Forcing; only the initialization differs.

7. [Demo Website] Full Video Comparisons (#fpMz-W2)

8. [Contribution & Limitations] Novelty and Extractor Decoupling (#oULG-W3/Q4)

  • Contribution: For few-step causal AR initialization, changing the regression metric suffices to replace both the intermediate teacher and the ODE stage, supported by the corollary and the extractor, depth and attention-mask ablations.
  • Limitations: Table 2 decouples the extractor from the fixed DMD teacher; a different-family extractor and training on videos from another generator are listed as future work.

We thank the reviewers again for their thoughtful feedback, which has substantially strengthened the paper, and welcome any further questions before the discussion period closes on November 18.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.