acceptodds
Under review as a conference paper at ICLR 2027

Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs

Abstract

Video generation models can now depict remarkably realistic motion. Yet it is unclear how to use this motion beyond pixels: to animate a specific 3D asset and replay the animation in its 3D scene.For instance, a video model can show a monkey jumping over a crossbar, but it cannot return that jump as replayable 3D motion placed in the 3D scene.Here we present a model that generates, together with the video, 3D motion for any supplied rig and a camera that registers this motion to the scene. Prior joint video–motion models copy billions of parameters from the video model to generate motion. We instead keep 3D motion in a separate model that supports arbitrary skeletons, and connect it to a pretrained video model through a small set of cross-attention layers. This separation lets the motion model first learn how skeletons move from motion-only data, which is easier to obtain and augment than paired video. Additionally, our model learns to locate each joint in the video frames, which lets us recover the camera and place the 3D motion in the scene. To train across many skeletons, we use a generative video editor to place animated 3D assets in realistic scenes, curating the edits so their known motion remains valid. Together with human and robot video, this yields nearly 200K training pairs. Our evaluations demonstrate that our model animates diverse rigs from text more faithfully than video-first pipelines, and its motion responds to the surrounding scene. It also turns video into rig motion, rig motion into video, and enables keyframe-base control of the 3D motion.

Primary area:
generative models
Keywords:
Video Generation, 3D Motion Generation, Articulated Motion, Arbitrary Skeletons, Joint Generation, Diffusion Modelstl, dr

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_H6r1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

JCA jointly generates video and 3D motion for arbitrary rigs, then estimates a camera to align the motion with the video. It connects a separately pretrained motion model to a video backbone through cross-attention bridges. The paper introduces a paired-data pipeline and bridge guidance, with evaluations on rig animation, human video–motion generation, and motion recovery.

Critical strengths
  1. The problem is useful: one model generates video and 3D motion for different rigs, with camera alignment for scene placement. It also supports video-to-motion, motion-to-video, and keyframe control.
  2. Bridge guidance is simple and effective. In Table 3, it reduces PA-MPJPE from about 85 to 72, while disconnecting the bridges increases the error to about 141.
  3. The data pipeline is practical. Video editing places animated assets in realistic scenes, and filtering checks label consistency. The resulting paired corpus covers a wide range of rigs.
  4. The results are encouraging: JCA achieves better prompt adherence than the tested rig-animation baselines, and the human-fine-tuned model outperforms EchoMotion on synchronization.
Critical weaknesses
  1. Although Puppeteer and ActionMesh are included, a direct Wan2.2 → TopoCap/MoCapAnything V2 baseline is missing. This makes it harder to judge the benefit over a strong video-generation-plus-motion-recovery pipeline.
  2. There is no matched video-quality comparison with the original Wan2.2. It is therefore unclear whether LoRA adaptation and joint training preserve the backbone's video quality.
  3. The human evaluation reports synchronization and video quality, but not standard text-to-motion metrics such as R-Precision, FID, MM-Dist, and Diversity. These would help assess text alignment and motion quality beyond agreement with the video.
Questions for authors
  1. Do you plan to publicly release the dataset, code, and model weights?
2 comments
AuthorsOpenReview ↗

Response to Reviewer H6r1

Dear Reviewer H6r1, thank you for supporting acceptance and for finding the problem "useful," bridge guidance "simple and effective," and the results "encouraging."

The main concern was evaluation coverage: direct video-first baselines, a matched video-quality comparison with Wan2.2, and standard text-to-motion metrics. We added all three. We summarize the new results first, then give details below.

  • JCA outperforms the requested video-first pipelines. Its prompt adherence is 58.4, versus 40.2 for Wan2.2 → MoCapAnything V2 and 24.0 for Wan2.2 → TopoCap, with comparable smoothness and mesh distortion.
  • Joint training largely preserves Wan2.2's video quality. VBench scores stay close to those of the original Wan2.2.
  • Our human-fine-tuned model achieves the best standard text-to-motion scores. It leads in R-Precision, FID and MM-Dist, for example R@1 of .305 vs. .272 and FID of 5.32 vs. 6.19 against EchoMotion.

Regarding release, we plan to publicly release the code, trained model checkpoints and data, subject to a formal licensing review. We believe this addresses the reviewer's concerns and question, and are happy to discuss further.

 

Answers to the concerns

Although Puppeteer and ActionMesh are included, a direct Wan2.2 → TopoCap/MoCapAnything V2 baseline is missing.

Thank you for the suggestion. We added both direct baselines, Wan2.2 → TopoCap and Wan2.2 → MoCapAnything V2, to the revised Table 1. JCA achieves substantially higher prompt adherence: 58.4, compared with 24.0 for TopoCap and 40.2 for MoCapAnything V2. MoCapAnything V2 achieves higher smoothness (.991 vs. .983) and lower mesh distortion (.16% vs. .20%), but substantially lower dynamic degree (.45 vs. .83).

MethodPrompt adherenceMotion SmoothnessDynamic degreeMesh distortion(%)
Wan2.2 → TopoCap24.0.981.88.63
Wan2.2 → MocapAnything V240.2.991.45.16
JCA58.4.983.83.20

 

There is no matched video-quality comparison with the original Wan2.2. It is therefore unclear whether LoRA adaptation and joint training preserve the backbone's video quality.

Following the reviewer’s suggestion, we expanded the video-quality evaluation from Table 2 with additional VBench metrics and the original Wan2.2 baseline, reporting the new results in App. I, Table 10. Our model achieves comparable video quality overall, with improvements on some metrics and small decreases on others, suggesting that LoRA adaptation and joint training largely preserve the backbone’s video quality.

MethodIQ ↑Aesth. ↑Smooth. ↑Anatomy ↑Subj. consistency ↑Bg. consistency ↑ViCLIP ↑Flicker ↑Dynamic degreeNon-static camera (%)
Original Wan2.2.673.598.993.936.9750.9515.1546.9949.46865.6
JCA.664.578.992.934.9516.9441.1556.9952.60045.6
JCA SMPL-FT.657.573.994.938.9591.9547.1559.9964.48844.8

 

The human evaluation reports synchronization and video quality, but not standard text-to-motion metrics such as R-Precision, FID, MM-Dist, and Diversity. These would help assess text alignment and motion quality beyond agreement with the video.

To address the reviewer’s request for standard text-to-motion evaluation, we extended the evaluation in Table 2 with R-Precision, FID, MM-Dist, and Diversity on the same HumanML3D test prompts from Section 4.3, with new results in Table 9 and the evaluation protocol in revised App. G. These results mirror the synchronization results in Table 2: the general JCA model trails the human-specific EchoMotion model, while JCA SMPL-FT substantially outperforms it. It achieves the highest R-Precision at all reported ranks and the lowest FID and MM-Dist among the compared methods. Relative to EchoMotion, R@1 improves from 0.272 to 0.305 and FID decreases from 6.185 to 5.324, supporting text alignment and motion quality beyond video–motion synchronization.

MethodR@1R@2R@3FIDMM-DistDiversity
EchoMotion.272.424.5366.1855.0527.122
JCA.246.404.52511.9275.3156.689
JCA SMPL-FT.305.482.5985.3244.7627.844

 

Do you plan to publicly release the dataset, code, and model weights?

Yes. We plan to publicly release the code, trained model checkpoints and data. Because our models and the OiV dataset are built from third-party assets, videos and motion data whose licenses differ, their release is subject to a formal licensing review.

Thanks for the detailed rebuttal. This answers my questions.

The video-first baselines, the Wan2.2 quality comparison, and the text-to-motion metrics are all what I asked for, and the results look convincing to me.

No remaining concerns from me. I'm keeping my score at 3 (weak accept).

Looking forward to the code and model release!

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

JCA couples an independently pretrained, rig-conditioned motion generator with a pretrained video diffusion model through bidirectional attention and bridge guidance. A learned 2D joint-localization head enables static-camera fitting. The authors also construct a large paired video–motion corpus through generative editing and demonstrate joint generation, conditional generation, and keyframe control. Evaluation includes text-driven rig animation, human video–motion synchronization, conditional animal motion recovery, and qualitative scene-aware animation.

Critical strengths
  1. A useful, flexible architecture. Separating the motion model permits motion-only pretraining and variable skeletal structures without duplicating a large video backbone for motion. This is a meaningful design for reusing heterogeneous motion data.

  2. Positive evidence for cross-modal coupling. Table 3 supports the benefit of the bridge connections and guidance; the motion-only comparison also shows improved prompt adherence after joint training. Supporting multiple conditional directions within one system adds practical value.

Critical weaknesses
  1. The defining arbitrary-rig co-generation claim lacks direct quantitative validation. Table 1 evaluates animation quality, Tables 2–3 evaluate human synchronization, and Table 4 evaluates conditional animal motion recovery. None directly measures whether jointly generated non-human video and 3D motion agree in pose and timing. Individually plausible outputs need not be mutually consistent, making this a consequential gap in the central evaluation.

  2. Scene registration and geometric responsiveness remain insufficiently validated. The synchronization metrics in Appendix C remove global translation through per-frame alignment or root centering, so they cannot assess global translational placement. Appendix A.5 fits a camera to predicted correspondences without changing motion; this alone does not establish correct depth or obstacle relationships in the input scene. The scene re-renderings provide useful qualitative evidence, but the broader scene-related claims lack corresponding systematic evaluation. I am not asking for a general physical-consistency guarantee.

These gaps concern the paper's distinguishing claims. Direct evidence or an adequately justified narrowing of scope could address them.

Questions for authors
  1. Direct validation: Can you evaluate video–motion pose and temporal agreement on a representative subset of existing jointly generated non-human samples, with an appropriate correspondence protocol or blinded assessment?

  2. Scope clarification: Does registration claim only image-space alignment, or placement consistent with the original 3D scene? What evidence distinguishes them? A small controlled scene-variation test would help assess geometric responsiveness.

  3. Decoding clarification: How do predicted rotations enter the final IK/export pipeline? Is the comparison with MoCapAnything V2 restricted to position metrics, or can the final rotations also be evaluated under matched conventions?

  4. Baseline clarification: Why is Wan → TopoCap absent from Table 1 despite TopoCap being used in Appendix D? A matched comparison, if compatible, would clarify robustness to the chosen recovery method.

  5. Reproducibility clarification: Please specify the joint-training schedule and actual bridge-guidance scales, provide the remaining training/inference configurations, and clarify the availability of data, trained checkpoints/adapters, and code, including any access restrictions.

8 comments
AuthorsOpenReview ↗

Response to Reviewer TFEB (1/2): Main Concerns

Dear Reviewer TFEB, thank you for the constructive feedback and for recognizing our design as "useful", "meaningful" and "practical".

Two concerns were raised: (i) whether the generated video and 3D motion agree for non-human rigs, and (ii) whether the motion's placement is consistent with the original 3D scene. New experiments address both. We summarize the new results first, then give details below.

  • Video and 3D motion match across rigs. Measured against joints tracked in the generated video, the 3D motion's error is 2.37%, versus 4.4% when the video and motion models are disconnected.
  • The motion responds to target placement in the 3D scene. In the controlled test the reviewer suggested, a character must reach a target placed at random. It ends 0.19 m away on average, having started 0.86 m away, while motions generated for other targets end 1.25 m away.

We answer the reviewer's additional questions in a separate response, including rotation accuracy and a video-first baseline (TopoCap).

We believe this addresses the reviewer's concerns and questions, and are happy to discuss further. The experiments are reported in the updated manuscript and supplement, and on a new supplementary webpage.

 

Answers to the concerns

Arbitrary-rig co-generation validation. Can you evaluate video–motion pose and temporal agreement on a representative subset of existing jointly generated non-human samples?

Measuring agreement for arbitrary rigs is challenging: for humans, Tables 2–3 rely on an established 3D pose estimator (GVHMR), but arbitrary rigs lack an equally reliable one. We address this by measuring agreement in image space, where joints of any rig can be tracked. TAPNext++ [1] tracks each 2D joint through the generated video from its first-frame location. We project the generated 3D joints with one camera per clip, fitted to best match these tracks, so the score does not depend on our own camera estimate. The metric measures projected pose and motion agreement after fitting one camera for the entire clip. Error is the mean joint distance relative to the image diagonal, excluding the given first frame.

On the 100 asset-prompt pairs of Table 1 (20 assets), error is 2.37% for full JCA, 2.46% without bridge guidance, and 4.40% with bridges disconnected (revised App. D, Table 6).

 

Scene-related claims lack corresponding systematic evaluation... (is) placement consistent with the original 3D scene? A small controlled scene-variation test would help assess geometric responsiveness.

We thank the reviewer for this suggestion. We conducted the proposed controlled test, detailed in the newly added Sec. 5 of the main paper and illustrated on the new supplementary webpage. In this test, a character must reach a floor target at 100 random locations in a 3D scene. Rendered back into the scene, its generated motion ends 0.185 m from the target, having started 0.856 m away. This demonstrates target-responsive placement in the original 3D scene, not only in the image.

Setup. Only the target's position varies; all else is fixed. First and last RGB keyframes specify the intended interaction. As 2D images, they specify the target only in the image, so the motion's 3D placement, including depth, is left to the model and registration. We fit the camera's rotation and translation to the predicted 2D joints, using the rendered frame's focal length, and measure the character's ground-plane distance to the target disk.

As a control, we score each target against a motion generated for a different target (shuffled pairs). To isolate the role of registration, we also use two simpler cameras: one fitted to the first frame only, and the render camera without fitting (no alignment).

Final distance (m)Full-sequence fit (ours)First-frame fitNo alignment
Assigned target0.1850.5750.571
Shuffled pairs1.2451.4591.539

Results. In every configuration, motions end closer to their assigned target than they started (0.856 m) and far from shuffled targets. This holds even without alignment, where all trials share one camera, so the generated motion itself responds to its target. Registration then cuts the remaining distance about threefold. For qualitative videos, see our supplementary webpage or the attached supplementary material.

Grounding in the scene. Placement draws on the video model's implicit 3D prior, the motion model's explicit 3D skeleton, and calibration to the scene's camera. Obstacle relationships, such as the crossbar in Fig. 1, are shown qualitatively. We agree with the reviewer that a general physical-consistency guarantee is beyond what this work aims to provide, as discussed in Sec. 7.

 

[1] Jung et al., TAPNext++, CVPR Findings 2026

AuthorsOpenReview ↗

Response to Reviewer TFEB (2/2): Additional Questions

Answers to questions

 

Direct validation: Can you evaluate video–motion pose and temporal agreement on a representative subset of existing jointly generated non-human samples, with an appropriate correspondence protocol or blinded assessment?

This is addressed in our response to the arbitrary-rig co-generation concern above.

 

Does registration claim only image-space alignment, or placement consistent with the original 3D scene? What evidence distinguishes them? A small controlled scene-variation test would help assess geometric responsiveness.

This is addressed in our response to the scene-related concern above.

 

Decoding clarification: How do predicted rotations enter the final IK/export pipeline? Is the comparison with MoCapAnything V2 restricted to position metrics, or can the final rotations also be evaluated under matched conventions?

For animation export, we use the predicted root orientation and a root trajectory reconstructed from predicted velocities to transform the predicted joint positions into world coordinates; inverse kinematics (IK) then recovers the exported rotations. This position-based reconstruction follows AnyTop and MoMask, and we found it preferable in preliminary qualitative joint-generation experiments. Following the reviewer’s request, we additionally evaluate local rotation errors on the TrueBones Zoo unseen test split used in App. E, under matched rotation conventions. Ours-FT obtains rotation errors of 22.74° for direct predictions and 32.64° for IK-exported rotations, compared with 19.56° for MoCapAnything V2’s output, a dedicated motion-capture model.

 

Baseline clarification: Why is Wan → TopoCap absent from Table 1 despite TopoCap being used in Appendix D? A matched comparison, if compatible, would clarify robustness to the chosen recovery method.

Following the reviewer’s suggestion, we added baselines applying TopoCap (and MoCapAnything V2) to Wan2.2-generated videos in the revised Table 1. Compared with the TopoCap baseline, our method achieves higher prompt adherence (58.4 vs. 24.0), smoother motion, and lower mesh distortion, with slightly lower dynamic degree.

 

Reproducibility clarification: Please specify the joint-training schedule and actual bridge-guidance scales, provide the remaining training/inference configurations, and clarify the availability of data, trained checkpoints/adapters, and code, including any access restrictions.

We supplement the details in App. A.7. with the joint-training schedule, full optimizer configuration, and inference settings. Specifically, joint training uses batch size 32 and learning rate 10⁻⁴ for 50K updates, followed by 20K updates at 10⁻⁵. Inference uses UniPC with 50 denoising steps and flow shift 3; text guidance is 3 and bridge guidance is 2 for both modalities. Videos are generated at 560 × 320, with 81 frames at 16 FPS.

We plan to publicly release the code, trained model checkpoints, including the video LoRA adapters, and data. Because our models and the OiV dataset are built from third-party assets, videos and motion data whose licenses differ, their release is subject to a formal licensing review.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer TFEB,

Before the discussion closes, we would appreciate hearing whether our new experiments resolve your main concerns.

Following your suggestions, we added the controlled scene-variation test (0.19 m error versus 1.25 m for shuffled pairs), measured non-human video–motion agreement (2.37% versus 4.40% error), and addressed your additional questions.

Your suggestions have helped us strengthen the paper. We would welcome your updated assessment, including any remaining concerns.

Many thanks,

The authors

Reviewer_F4FuOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents JCA, a framework that jointly generates videos and synchronized 3D motion for arbitrary skeletal rigs from text and optional reference images. It couples an independently pretrained motion model with a pretrained video diffusion model through bidirectional cross-attention. Bridge guidance strengthens cross-modal consistency, while learned 2D joint localization enables camera estimation and motion alignment. Trained on approximately 200,000 video–motion pairs, JCA supports joint generation, video-to-motion recovery, motion-to-video synthesis, and image-keyframe control.

Critical strengths

The paper presents a practical framework for jointly generating videos and reusable 3D motion across diverse skeletal rigs. Its independently pretrainable motion model, parameter-efficient bidirectional coupling, and joint localization for camera alignment are notable strengths. The unified architecture also supports motion recovery, motion-controlled video generation, and keyframe conditioning. In my opinion, this contributes somewhat for 3D content generation.

Critical weaknesses

Main limitation of this paper lie in its infficient ablation experiments:

  1. Experiments only demonstrate that bidirectional video–motion coupling is better than motion-only module in test (used pretrained motion DiT). I think the author takes a step further to check how motion DiT and video DiT contribute each other, and how their coupling affects the performance. So they should add more abalation studies to several variants, including video only, motion-to-video only, video-to-motion only etc.
  2. Joint localization and camera alignment is another key-point of this paper, but no further experiments to demonstrate how it helps the model's performance. I want to know how the model performs when there is no alignment, first-frame fitting, and full-sequence fitting using predicted 2D joints.
Questions for authors

see Critical Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer F4Fu

Dear Reviewer F4Fu, thank you for the constructive feedback and for recognizing JCA as "a practical framework" for "jointly generating videos and reusable 3D motion across diverse skeletal rigs," with "notable strengths" in its design.

New experiments address both main requested ablations: (i) how the video and motion models contribute to each other, and (ii) how much camera alignment helps. We summarize the new results first, then give details below.

  • Joint generation outperforms separate or sequential generation. Jointly generated motion follows the prompt better than motion generated blind to the video or after it (58.4 vs. 44.9 and 53.6). Outputs agree more closely than when disconnected (2.37% vs. 4.40% error), with comparable video quality. Each model also predicts better when connected.
  • Camera alignment is essential for placement. It roughly halves image alignment error (3.1% vs. 5.5-5.9%) and brings the motion three times closer to its target in a 3D scene (0.19 m vs. 0.57 m).

We believe this addresses the reviewer's concerns, and are happy to discuss further. The experiments are reported in the updated manuscript and supplement, and on a new supplementary webpage.

 

Answers to the concerns

check how motion DiT and video DiT contribute each other... add more ablation studies including video only, motion-to-video only, video-to-motion only etc.

We add ablations and two direct coupling measurements showing that each branch benefits from the other. Motion makes the video controllable with comparable quality, while coupling lowers both branches’ denoising error and improves video–motion agreement. Video helps motion follow the prompt: on Table 1’s asset-prompt pairs, joint generation reaches 58.4 prompt adherence, versus 53.6 for Wan2.2 video followed by our video-to-motion mode and 44.9 when motion cannot attend to video at inference (revised App. B.4, Table 4)

VariantPrompt ↑Smooth. ↑Dyn.Distort. (%) ↓
Wan2.2 video, then our video-to-motion53.6.978.910.23
Joint, motion cannot attend video44.9.986.570.15
Joint, full model58.4.983.830.20

Motion controls the video, with comparable quality. On the same pairs, our VBench scores are on par with off-the-shelf Wan2.2. Coupling adds control: motion-to-video follows a given 3D trajectory across rigs (Sec. 4.4, Figs. 1, 4). Motion-first scores higher on consistency, but its generated videos are less dynamic (48% vs. 78%; App. B.4, Table 4b).

VariantIQAesth.Smooth.Dyn. %Subj.Bg.TextI2V subj.I2V bg.
Joint, full model.480.588.99278.898.934.194.940.949
Joint, video cannot attend motion.477.583.99274.899.934.191.947.954
Joint, bridges off.478.583.99274.899.935.191.947.955
Motion first, then motion-to-video.488.600.99348.921.944.189.955.958
Off-the-shelf Wan2.2.477.559.98776.878.921.197.918.931

Each branch predicts better with the other. On OiV-test, a held-out subset of 200 clips from OiV, we compare one-step denoising error with bridges on and off. Bridges lower both branches' error at every noise level, most in early high-noise steps that set structure (detailed in App. B.4, Table 5). Training also drops the bridges, so both states are in-distribution.

The outputs agree. On a new tracking-based metric (see reply to TFEB), video-motion error is 2.37% for full JCA, 2.46% without bridge guidance, and 4.40% with bridges disconnected (revised App. D, Table 6).

 

localization and camera alignment... how the model performs when there is no alignment, first-frame fitting, and full-sequence fitting using predicted 2D joints.

Following the reviewer's suggestion, we added two experiments evaluating camera alignment in 2D and 3D. Full-sequence fitting outperforms both alternatives: it reduces 2D alignment error by 44-47% and makes 3D placement about three times more accurate.

2D alignment. On OiV-test, we fix videos and motions and vary only the projection camera. No alignment keeps the camera that rendered the input frame. First-frame fitting fits it to the input first-frame 2D joints. Full-sequence fitting (ours) fits it to the predicted 2D joints of all frames. The mean distance between projected joints and TAPNext++ [1] tracks in the generated video, relative to the image diagonal, is 5.85%, 5.53% and 3.08%, respectively (revised App. H, Table 7).

Placement in a 3D scene. In a controlled test reported in the newly added Sec. 5 of the main paper (see reply to TFEB), a character must reach a floor target at 100 random locations. On average, the motion ends 0.185 m from the target with full-sequence fitting, versus 0.575 m and 0.571 m with first-frame fitting and no alignment.

 

[1] Jung et al., TAPNext++, CVPR Findings 2026

Thanks for your new experiments, which answer my concerns on motion DiT and video DiT's role. Also 2D alignment and placement in a 3D scene experimental results reveal that full-sequence fitting is important in your model. I will improve my score from weak rejection to weak acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank all reviewers for their careful reading and constructive feedback. We appreciate their recognition of JCA’s flexible architecture for arbitrary rigs [F4Fu, TFEB, H6r1], the value of independently pretraining the motion model [F4Fu, TFEB], and the effectiveness of video–motion coupling [TFEB, H6r1]. The reviews identify useful ways to strengthen the experimental evaluation and clarify the method and the scope of our claims.

The main questions concerned 3D scene placement [F4Fu, TFEB], video–motion agreement across arbitrary rigs [F4Fu, TFEB], and additional quality evaluations [TFEB, H6r1]. We believe our new experiments and ablations provide evidence addressing these concerns. Detailed responses are provided in the individual review threads. The results and clarifications are presented in the revised manuscript and supplement, with changes and additions highlighted in blue. Additional visual results are available on the added supplementary webpage.

We thank the reviewers again for their time and suggestions. We believe these additions strengthen the paper and welcome further discussion to address any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.