acceptodds
Under review as a conference paper at ICLR 2027

CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion

Abstract

Diffusion-based video generation models (VGMs) can produce visually plausible videos, yet their camera ego-motion often drifts over long horizons. Camera ego-motion determines how the viewpoint changes between frames and is essential for maintaining consistent geometry and trajectories. To address this problem, we propose CausalAlign, a causal alignment framework with two components. First, we diagnose how precise ego-motion is represented, used, and retained within VGMs using linear probes and causal interventions, revealing a rise-and-fall pattern during denoising. Second, we construct a delta-token space from a 3D foundation model and align VGM representations at causally identified positions. Delta tokens encode changes between consecutive frames, allowing the alignment to focus on ego-motion rather than static scene content. With Wan2.2 fine-tuned on DL3DV and RealEstate10K, CausalAlign preserves visual quality while improving motion accuracy, geometric consistency, and temporal smoothness. For example, it reduces geometric consistency errors by more than 20% relative to the base model. Moreover, using the fine-tuned VGM, CausalAlign improves downstream world action models (WAMs) on the RoboCasa365 benchmark. Overall, CausalAlign improves ego-motion consistency in diffusion-based VGMs and yields gains in downstream embodied models on motion-intensive tasks.

Primary area:
generative models
Keywords:
Video Generation, Ego-Motion Consistency, Representation Alignment, Causal Interventions, World Action Models

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_UpPYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CausalAlign, a framework for improving camera ego-motion consistency in video generation models. The authors analyze internal ego-motion representations using linear probing and causal interventions, and identify a rise-and-fall pattern during denoising, termed geometric state forgetting. Based on this observation, CausalAlign aligns VGM representations with a delta-token space learned from a 3D foundation model at causally selected layer–timestep locations. Experiments on Wan2.2-14B show improved camera motion, geometric consistency, and temporal dynamics while largely preserving visual quality, with additional gains on downstream RoboCasa365 tasks.

Critical strengths
  1. The paper provides a systematic analysis of ego-motion representations inside VGMs rather than relying only on geometry metrics of generated videos. By combining linear probing with causal intervention, it distinguishes whether ego-motion information is decodable from hidden states from whether it actually influences the generation process, and further identifies the geometric state forgetting phenomenon.

  2. The method is well motivated by the preceding analysis. Instead of aligning 3D features at arbitrarily chosen layers, CausalAlign targets locations where ego-motion information begins to decline, and uses delta tokens to emphasize inter-frame motion and geometric changes rather than static scene content.

  3. The experiments provide fairly comprehensive evidence for the method. CausalAlign improves camera-motion consistency, geometric consistency, and temporal dynamics while largely maintaining visual quality. Its transfer to Fast-WAM also suggests that improving geometric representations in the VGM can benefit downstream robot-control tasks.

Critical weaknesses
  1. The ablations do not fully disentangle the contributions of the different components in the complete training recipe. Although the paper includes a matched layer-placement study on the 14B model, it would be useful to further compare against all-layer LoRA fine-tuning under a matched budget, delta-token alignment without the explicit pose loss, and random or uniformly selected layer–timestep locations using the same losses. Such controls would clarify how much of the gain comes from causal target selection, delta-token representation, and pose supervision, respectively.
Questions for authors
  1. The main analysis of geometric state forgetting is conducted on Wan2.2-14B. Have the authors repeated the causal diagnosis on other model backbones or model scales and observed a similar rise-and-fall pattern? This would help establish whether the phenomenon is general across video generation models and strengthen the claimed generality of the method.
3 comments
AuthorsOpenReview ↗

Official Comment to Reviewer UpPY from Authors

Thank you for your careful review. We now respond to your questions as follows.

W1 (part 1). Additional ablation study on losses and token structure.

The short conclusion is that combining pose supervision and delta matching achieves the best results, and structured delta tokens performs best among the tested alignment targets. As requested by the reviewer, we compare the full objective with pose-only and delta-only training under the same settings. In Table 10, combining delta matching loss and pose supervision loss improves all five metrics(e.g.,reducing reprojection error from 6.9 to 6.5). Furthermore, We also change the alignment target while keeping pose supervision fixed. In Table 11, structured delta tokens yield lower geometric errors and jerk than unpartitioned 128D alternatives. These results show that delta matching loss adds value beyond pose supervision and that structuring the tokens improves the effectiveness of alignment.

W1 (part 2). Additional ablation study on diagnosed layers.

The short conclusion is that diagnosed layers achieve the lowest geometric errors among all tested layer sets. As requested by the reviewer, we compare them with random, evenly spaced, and all-layer selections while holding the timesteps, temporal weights, and losses fixed. For the all-layer control, we reduce the per-layer LoRA rank to keep the total number of trainable adapter parameters comparable. In Table 13, joint selection for diagnosed layers outperforms all other sets (e.g., reprojection error compared with all-layer LoRA, 6.5 px vs. 8.9 px). This comparison supports our design of using causal diagnosis to choose which layers to adapt.

W1 (part 3). Additional ablation study on diagnosed timesteps.

The short conclusion is that the diagnosed timesteps achieve the best performance among all tested timestep sets. As requested by the reviewer, we compare the diagnosed timesteps with random and evenly spaced selections while holding the layers and losses fixed. We first use uniform weights for all selections to isolate the effect of timestep choice. In Table 13, the diagnosed timesteps outperforms all other sets (e.g., reprojection error compared with evenly spaced timesteps, 6.5 px vs. 8.0 px). These results support using causal diagnosis to select the timesteps.

Q1. Does the pattern appear on a second backbone?

The rise-and-fall pattern can be also observed at a different model family with a different model size. As requested by the reviewer, we again perform the linear probing and causal interventions on the frozen CogVideoX-5B-I2V which is another open-source, widely-used model. Figure 6 shows that decodability and intervention effects both rise and then fall across layers and timesteps, which has the same trend of our main experiments. So this pattern might be general in video generation models and a study on more generation models would be interesting in the future work.

If these experiments and clarifications have addressed your concerns, we would be grateful if you could consider raising your rating.

Thank you for your detailed response and additional experiments. I believe the rebuttal has effectively addressed my concerns. Therefore, I am happy to maintain my original positive rating.

Thank you for considering the new results and for sharing your positive assessment of the revised work.

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

Summary: Useful Causal Diagnosis, but Limited Novelty and Weak Attribution of the Gains

This paper studies camera ego-motion drift in video generation models. It uses linear probes to test whether relative camera pose can be decoded from hidden states, and Mask, Shuffle, Replace, and Inject interventions to test whether those states affect generated trajectories. Both signals rise and then fall across Transformer layers and denoising progress, which the paper calls “geometric state forgetting.”

CausalAlign learns a 128-dimensional delta-token space from LingBot-Map, maps VGM frame-to-frame features into that space, and trains LoRA at selected layers and denoising positions with delta matching and relative-pose losses. Only the adapted VGM is retained at inference.

Critical strengths
  1. The diagnosis distinguishes information that is readable from information that affects generation, and the added controls make the rise-and-fall result more credible.
  2. The 3D teacher is used only during training, and the video and RoboCasa365 experiments give initial downstream evidence.
Critical weaknesses
  1. The novelty mainly lies in combining existing ideas. The paper uses AC3D as a premise for its analysis, describes Geometry Forcing as direct feature alignment without causally selected layer-timestep positions, and states that the delta-token idea comes from DeltaTok [2,4,5]. However, REPA already established frozen-teacher alignment for diffusion features, and VideoREPA extended representation alignment to video generation [1,3]. CausalAlign combines target selection based on the causal diagnosis, structured delta tokens, and LoRA at selected layers, but the experiments do not separate their individual contributions.
  2. Table 4 tests how well pose and geometry can be read from delta tokens; it does not show that these tokens improve generation, and the tokens receive explicit pose supervision. The paper lacks a same-backbone, same-data, and same-budget comparison with direct feature alignment, direct pose supervision, frame-feature differences, and random or all-layer alignment. Table 2 matches output and evaluation settings, but some input interfaces differ and matched backbones or training budgets are not established. Mean Translation/Angular Proximity also measures only proximity to the real mean motion magnitude, not the full motion distribution.

[1] Yu et al., “Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think,” ICLR 2025 Oral: https://arxiv.org/abs/2410.06940
[2] Bahmani et al., “AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers,” CVPR 2025: https://arxiv.org/abs/2411.18673
[3] Zhang et al., “VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models,” NeurIPS 2025: https://doi.org/10.52202/085713-4091
[4] Wu et al., “Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modeling,” ICLR 2026: https://arxiv.org/abs/2507.07982
[5] Kerssies et al., “A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens,” CVPR 2026: https://arxiv.org/abs/2604.04913

Questions for authors
  1. Under the same Wan2.2-14B backbone, data, and parameter budget, how do direct 3D feature alignment, frame-feature differences, structured delta tokens, and full CausalAlign compare?
  2. What are the results after separately removing LposeL_{\text{pose}} and LΔL_\Delta, or using random layers, a fixed set of middle layers, all layers, and all timesteps?
  3. Can the Fast-WAM study add a matched control without the CausalAlign losses and target selection, with multiple training seeds or confidence intervals?
3 comments
AuthorsOpenReview ↗

Official Comment to Reviewer wqBr from Authors (1/2)

Thank you for your careful review and for pointing us to relevant works. We have added these references and clarified how they relate to CausalAlign.

W1. Novelty and relationship to prior work.

Both our work and AC3D study camera motion in video generation. However, the motivation, task setting, modeling design, diagnosis method, and the conclusion are all different. AC3D aims to generate a video guided by a given camera trajectory, whereas we aim to improve ego-motion consistency when no trajectory is provided at inference. Accordingly, AC3D uses spectral analysis and pose probes to identify when camera motion emerges and where it is decodable. We instead perform pose probing and the hidden-state interventions to measure both decodability and influence on the generated trajectory. AC3D observes that camera motion emerges early in denoising and uses this finding to restrict camera-trajectory conditioning to early steps. CausalAlign pairs pose probes with hidden-state interventions to measure both how well ego-motion can be decoded and how strongly the hidden states affect the generated trajectory. Both signals decline after peaking. We use these declines to select the layers and timesteps at which to apply geometric supervision during fine-tuning.

DeltaTok is designed for a better efficiency of world model, while we adapt the high-level idea of delta token to provide a better supervision for geometry correction. The similar high-level idea is used in predictive video compression, such as DVC (Lu et al., CVPR 2019). REPA, VideoREPA, and Geometry Forcing establish teacher alignment, relational alignment for video, and alignment to 3D features, respectively. In contrast, our method aligns the delta token space derived from a 3D foundation model. Further, new experiments in Table 11 tests the value of our method against the learned, unpartitioned targets with the same 128-dimensional width and teacher-side supervision. Compared with the learned frame-difference target, structured delta tokens reduce reprojection error from 8.1 to 6.5 px.

Overall, our method has the high-level idea common in the community (i.e., the motion difference, representation alignment, 3D foundation model as a teacher). But our work is unique in diagnosing the geometry pattern in video generation models, designing a fine-tuning method to improve geometry states, and achieving strong results across video generation tasks and embodied tasks. We thus believe our work deserves an opportunity so that we can convey these messages to the community.

W2/Q1/Q2. Additional ablation study on losses and token structure.

The short conclusion is that combining pose supervision and delta matching achieves the best results, and structured delta tokens performs best among the tested alignment targets. As requested by the reviewer, we compare the full objective with pose-only and delta-only training under the same settings. In Table 10, combining delta matching loss and pose supervision loss improves all five metrics (e.g.,reducing reprojection error from 6.9 to 6.5). Furthermore, We also change the alignment target while keeping pose supervision fixed. In Table 11, structured delta tokens yield lower geometric errors and jerk than unpartitioned 128D alternatives. These results show that delta matching loss adds value beyond pose supervision and that structuring the tokens improves the effectiveness of alignment.

Q2. Additional ablation study on diagnosed layers and timesteps.

For Transformer layers, the short conclusion is that the diagnosed layers achieve the lowest geometric errors among all tested layer sets. As requested by the reviewer, we compare them with random, evenly spaced, and all-layer selections while holding the timesteps, temporal weights, and losses fixed. For the all-layer control, we reduce the per-layer LoRA rank to keep the total number of trainable adapter parameters comparable. In Table 13, joint selection for diagnosed layers outperforms all other sets (e.g., reprojection error compared with all-layer LoRA, 6.5 px vs. 8.9 px). This comparison supports our design of using causal diagnosis to choose which layers to adapt.

For denoising timesteps, the short conclusion is that the diagnosed timesteps achieve the best performance among all tested timestep sets. As requested by the reviewer, we compare the diagnosed timesteps with random and evenly spaced selections while holding the layers and losses fixed. We first use uniform weights for all selections to isolate the effect of timestep choice. In Table 13, the diagnosed timesteps outperforms all other sets(e.g., reprojection error compared with evenly spaced timesteps, 6.5 px vs. 8.0 px). These results support using causal diagnosis to select the timesteps.

AuthorsOpenReview ↗

Official Comment to Reviewer wqBr from Authors (2/2)

W2. What does the cross-method comparison establish?

Table 2 compares the motion and geometric consistency of videos produced by the full systems under their native settings. As clarified in Table 2 and Appendix B.2, some methods receive a camera trajectory or a 3D cache. These inputs can affect the videos when a method generates them. Each generated video is scored with the same output specifications, frozen estimators, and metric definitions, using a common real-video reference wherever required. The scores therefore measure the same output properties across methods, even when the systems use different inputs. To attribute the gains to individual design choices, Appendix C provides controlled comparisons that fix the Wan2.2-14B-I2V backbone and training data and match adaptation budgets as closely as each control allows.

W2. Do the motion gains extend beyond the mean?

Our work brings the distribution of generated camera speeds closer to that of real videos. As requested by the reviewer, we added Wasserstein-1 comparisons for translational and angular speed at the frame-pair and clip-mean levels. In Table 6, the frame-pair translation distance on DL3DV falls from 0.082 to 0.066. All eight distance metrics across two datasets are improved. These results explain the improvement in the measured speed distributions in addition to the mean proximity scores.

Q3. Does alignment account for the Fast-WAM gain?

As requested by the reviewer, we conduct additional ablation studies to Fast-WAM. In Appendix C.4, we compare our fine-tuning method with a video expert model fine-tuned by flow matching, using the same targeted layers, video data, and LoRA parameter budget. With both models ready, we adapt them to downstream embodied tasks with the same training and evaluation strategy. The results are shown in Table 15. Our method can largely outperforms the model fine-tuned by flow matching (i.e., 78.1% vs. 65.2%). Furthermore, we alse add a matched control with the video expert model fine-tuned by delta-token alignment at uniform layers. Our method still can largely outperforms it(i.e., 78.1% vs. 62.0%). These results validate the benefits of our proposed method on embodied tasks.

If these experiments and clarifications have addressed your concerns, we would be grateful if you could consider raising your rating.

Dear Reviewer,

We understand this is a busy period. With the discussion closing soon, given the concerns you raised in your initial review, we would be grateful to hear whether our rebuttal and the discussion with the other reviewers have helped address them, and whether the additional evidence might change your overall assessment. Thank you for your time.

Reviewer_Zd1SOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CausalAlign, a framework for improving ego-motion consistency in video generation models. The authors first use linear probes and causal interventions to identify a decline in ego-motion decodability and causal influence at later stages of denoising. They interpret this tendency as geometric state forgetting, and construct a motion-centric delta-token space using a pretrained 3D foundation model to address this problem. They finetune the VGM with LoRA to align its hidden representations with delta tokens at selected layers and timesteps identified by the diagnostic analysis. Experiments show improvements in ego-motion consistency and temporal dynamics while preserving visual quality, as well as gains in downstream world action model performance.

Critical strengths
  1. The paper provides a systematic analysis of ego-motion representations across transformer layers and denoising timesteps, using linear probing and causal interventions, to examine their role in video generation.
  2. The proposed method improves geometric consistency and temporal dynamics without directly optimizing smoothness or jerk in the final VGM finetuning objective, suggesting that the learned geometric representations may also benefit the temporal stability of generated videos.
Critical weaknesses
  1. The interpretation of geometric state forgetting is not sufficiently justified. The decline in decodability and causal influence may reflect a change in representation or the fact that camera geometry has already been established at earlier stages, rather than information loss. As suggested by AC3D, later layers may build upon camera information established in earlier layers. Thus, improved late-stage decodability after CausalAlign could result from additional geometric supervision rather than preventing forgetting.
  2. The contributions of the proposed components are not sufficiently isolated. Prior works have already explored probing-based layer selection and geometry-aware feature alignment. Since decodability and causal intervention show similar trends, it is unclear whether causal intervention provides additional benefits over probing alone. Furthermore, in the downstream WAM evaluation, comparing against a video expert initialized from a VGM fine-tuned on the same data with a comparable training budget, but without the proposed alignment, would help isolate the benefit of CausalAlign from additional finetuning.
Questions for authors

I would appreciate the authors' responses to the concerns raised in the Critical Weaknesses section.

3 comments
AuthorsOpenReview ↗

Official Comment to Reviewer Zd1S from Authors

Thank you for your thoughtful review. We now respond to your questions as follows.

W1 (part 1). What does the late decline mean?

As suggested by AC3D, camera geometry can form early and guide later layers. Different from AC3D, we concern about what happens after that geometry has formed. We find that earlier motion information can still help the model refine the output during late denoising. Restoring access to that information can therefore mitigate the effects of geometric state forgetting and improve the final geometry performance.

To validate this claim, we start from the same late latent and restore the motion state cached earlier for the same clip. We compare this intervention with ordinary denoising and with norm-matched edits to shuffled motion and non-motion content. As shown in Table 5, restoring the same clip's motion reduces reprojection error from 8.6 to 7.6 px, an improvement of about 12% over ordinary denoising, and yields lower error than both edit controls. This comparison shows that the model can use the specific motion information established earlier to improve the final geometry. Furthermore, this result might also explain why solving geometric state forgetting improves ego-motion consistency: preserving access to earlier motion information can help the model maintain a consistent camera trajectory through late denoising.

W1 (part 2). Does pose supervision explain the smaller decline?

Pose supervision explains part of the improvement, and delta alignment reduces the decline further. As requested by the reviewer, we added a pose-only control and measured the change in decodability from middle to late denoising under the same probe protocol. Table 5 shows that the drop is 0.19 with pose-only training and 0.09 with full CausalAlign. Pose-only training also reduces the drop relative to native Wan2.2. This supports delta-token alignment has better effect compared to pose supervision alone.

W2 (part 1). Do interventions improve layer selection?

Intervention measurements help to select better layers than probing alone. As requested by the reviewer, we added a comparison of probe-only, intervention-only, and joint selection with the same layer budget, timesteps, timestep weights, and losses. Table 13 shows that adding intervention evidence lowers reprojection error from 6.9 to 6.5 px relative to probe-only selection, with the same VBench quality score. This supports the necessary of using diagnosed layers when choosing where to supervise.

W2 (part 2). Does alignment account for the Fast-WAM gain?

As requested by the reviewer, we conduct additional ablation studies to Fast-WAM. In Appendix C.4, we compare our fine-tuning method with a video expert model fine-tuned by flow matching, using the same targeted layers, video data, and LoRA parameter budget. With both models ready, we adapt them to downstream embodied tasks with the same training and evaluation strategy. The results are shown in Table 15. Our method can largely outperforms the model fine-tuned by flow matching (i.e., 78.1% vs. 65.2%). These results validate the benefits of our proposed method on embodied tasks.

If these experiments and clarifications have addressed your concerns, we would be grateful if you could consider raising your rating.

Thank you for your detailed responses and the additional experiments. The new results have addressed my main concerns. I will maintain my original score.

Thank you for reviewing the additional experiments and for confirming that they addressed your main concerns.

Reviewer_N7yAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CausalAlign studies camera ego-motion in video generation models. Probes and hidden-state interventions show that motion signals rise at intermediate stages and then weaken. The method learns delta tokens from consecutive states of a frozen 3D teacher. It trains projection heads to map video features into this space. The heads are then frozen, and LoRA adapters are trained at selected layers using pose and token losses.Experiments cover I2V, T2V, camera-conditioned generation, and Fast-WAM transfer. Reported geometry and motion metrics improve. Visual quality stays close to the base model in the main I2V experiment. Ablations examine the token representation and layer placement.

Critical strengths

1.Internal diagnosis guides geometry supervision. The method uses probes and interventions to choose where to apply supervision. This gives a clear basis for layer selection. Delta tokens focus the supervision on changes between frames. The design requires no 3D teacher at inference.
2.The experiments support the method’s value. Tests across several generation settings show better camera-motion and geometric consistency. Visual quality stays close to the base model in the main I2V evaluation. Controlled comparisons also support the delta representation and the choice of layers.

Critical weaknesses

1.The evidence for harmful forgetting remains limited. Geometry commitment stays high in Fig. 5(c). Geometry may already be stored in the video latent, while later layers handle other computations. The scores in Eqs. (3) and (17) also depend on error relative to a fixed reference trajectory. That trajectory is not given to the generator. These scores therefore have limits as measures of the causal use of motion information.
2.The contribution of delta alignment to video generation needs clearer evidence. Table 4 tests feature readout accuracy. It does not show how much delta alignment adds beyond pose supervision or direct teacher-feature alignment.

Questions for authors

1.Could the late decline in motion signals reflect geometry already stored in the video latent? Please explain what Fig. 5(c) establishes. Does the evidence support harmful forgetting, or mainly reduced access to motion information?
2.What do the existing ablations show about the added value of delta alignment? Please distinguish the feature-readout results from the video-generation results.

3 comments
AuthorsOpenReview ↗

Official Comment to Reviewer N7yA from Authors

Thank you for your careful review. We now respond to your questions as follows.

W1 (part 1)/Q1. Does the late drop in motion signals matter?

The reviewer raises the possibility that the late decline in motion signals is harmless since the video latent already encodes the camera geometry. However, our hypothesis is opposite: the reduced access to earlier motion limits the model capability to refine camera geometry during late denoising. Therefore, restoring access to camera geometry can mitigate the effects of geometric state forgetting and improve the final geometry. We test this hypothesis by asking whether restoring the earlier geometry of a identical clip improves the final trajectory, when both feedforward start from the same late latent.

Specifically, in Table 5, restoring the earlier geometry reduces reprojection error from 8.6 to 7.6 px, about a 12% reduction relative to ordinary denoising, and outperforms both norm-matched controls using shuffled motion and non-motion content. This result helps explain why mitigating geometric state forgetting can improve ego-motion consistency. Appendix A.3 also defines geometry commitment K(s)K(s) as the minimum agreement across geometric information between the current clean prediction and the final video. The high K(s)K(s) in Fig. 5(c) suggests that much of the final geometry is already established early, and the recovery experiment shows that the model can still use earlier motion to refine it during late denoising. This also explains why mitigating geometric state forgetting improves ego-motion consistency: preserving access to earlier motion information helps the model refine the camera trajectory during late denoising.

W1 (part 2). Does the late decline depend on the reference trajectory?

The reviewer asks whether the late decline depends on scoring interventions against a fixed reference trajectory. To answer this question, we rescore the same intervention runs using the unperturbed generation from the same inputs and seed as the reference. At each denoising stage, we compute the trajectory displacement caused by the intervention and subtract the displacement caused by its norm-matched random control, both measured against that same unperturbed output. The results are shown in Table 5. The difference falls from 0.36 in the middle stage to 0.12 in the late stage. The decline thus persists when we measure changes relative to the model's own output.

W2/Q2. Does delta alignment improve the generated video?

As requested by the reviewer, we perform additional ablation studies on the loss or the alignment target, with the layers, timesteps and adaptation budget fixed. As shown in Table 10, we ablate on the loss design. The full loss (the last row) outperforms the pose supervision only (the third row) on all metrics (e.g., decreasing the reprojection error from 6.9 to 6.5 px). Their only difference is whether adding delta matching loss or not. So delta matching loss is beneficial and necessary. Furthermore, we ablate on different alignment targets in Table 11. Again, our structured delta target (last row) outperforms the other alignment targets on all metrics. Together, these new ablation results validate our design of loss and alignment target.

If these experiments and clarifications have addressed your concerns, we would be grateful if you could consider raising your rating.

Thank you for the detailed rebuttal and for providing the additional clarifications and paper details. They have addressed my questions well. I will maintain my original score.

Thank you for taking the time to read our rebuttal and for letting us know that the clarifications addressed your questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Comment to All Reviewers from Authors

Dear Reviewers and Area Chair,

Thank you for your careful reviews and constructive feedback. We have revised the manuscript in response to your comments. Before responding to each reviewer, we summarize the revision of our paper as follows.

Additional experiments and comparisons.

  • Additional ablations (all reviewers). The new ablations support combining pose supervision with delta matching, using structured delta tokens as alignment targets, and selecting layers and timesteps through causal diagnosis. The matched Fast-WAM controls also show that CausalAlign improves downstream success over flow-matching fine-tuning. We provide the details in the responses to each reviewer.

  • Late motion recovery (N7yA, Zd1S). Earlier motion can still improve the final video after much of the camera geometry has formed. Appendix A.3 defines geometry commitment K(s)K(s) by comparing the geometry of the current clean prediction with that of the final video from the same run. Starting from the same late latent, restoring the same clip's earlier motion reduces reprojection error from 8.6 to 7.6 px and outperforms both norm-matched edit controls. The late decline in intervention effects also persists when scored against the unperturbed output. The pose-only control shows a larger decodability drop than full CausalAlign (0.19 vs. 0.09) (Table 5).

  • Diagnosis on a second backbone (UpPY). On frozen CogVideoX-5B-I2V, decodability and causal intervention effects both rise and then fall across layers and denoising timesteps, as on Wan2.2 (Appendix A.4, Fig. 6).

  • Motion distributions (wqBr). CausalAlign reduces the Wasserstein-1 distance to real-video translational and angular speed distributions at the frame-pair and clip-mean levels. All eight distances improve over native Wan2.2 across DL3DV and RealEstate10K (Appendix B.1, Table 6).

Clarifications.

  • Novelty and relationship to prior work (wqBr). Our work performs both pose probes and the hidden-state interventions to choose which layers to align the representations, and leverages the structured delta tokens derived from a 3D teacher as the alignment target. Section 2 adds the discussions about REPA and VideoREPA and clarifies how CausalAlign relates to AC3D, Geometry Forcing, and DeltaTok.

  • Scope of the cross-method comparison (wqBr). Table 2 evaluates the full systems under their native settings. The conditioning inputs can affect the videos produced, and the same frozen estimators and metrics score the generated videos across systems. We clarify the input interfaces in Table 2 and Appendix B.2. The controlled comparisons in Appendix C fix the backbone and training data and match adaptation budgets as closely as each control permits.

  • Figure readability. Figure 1 separates the method overview from the results. Figure 2 makes the three training stages easier to distinguish.

Thank you again for your time and suggestions. We would be happy to address any remaining questions. If the new evidence and clarifications have addressed your concerns, we would be grateful if you could consider raising the score.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.