acceptodds
Under review as a conference paper at ICLR 2027

OmniHumanoid: Streaming Cross-Embodiment Video Generation with Paired-Free Adaptation

Abstract

Cross-embodiment video generation aims to transfer motions across different humanoid embodiments, such as human-to-robot and robot-to-robot, enabling scalable data generation for embodied intelligence. A major challenge in this setting is that motion dynamics are partly transferable across embodiments, whereas appearance and morphology remain embodiment-specific. Existing approaches often entangle these factors, and many require paired data for every target embodiment, which limits scalability to new robots. We present OmniHumanoid, a framework that factorizes transferable motion learning and embodiment-specific adaptation. Our method learns a shared motion transfer model from motion-aligned paired videos spanning multiple embodiments, while adapting to a new embodiment using only unpaired videos through lightweight embodiment-specific adapters. To reduce interference between motion transfer and embodiment adaptation, we further introduce a branch-isolated attention design that separates motion conditioning from embodiment-specific modulation. In addition, we construct a synthetic cross-embodiment dataset with motion-aligned paired videos rendered across diverse humanoid assets, scenes, and viewpoints. Experiments on both synthetic and real-world benchmarks show that OmniHumanoid achieves strong motion fidelity and embodiment consistency, while enabling scalable adaptation to unseen humanoid embodiments without retraining the shared motion model.

Primary area:
generative models
Keywords:
Video Generation, Robotic, Video Editting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Hh7ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper adapts the LoRA-based disentanglement paradigm explored in works such as OmniConsistency to cross-embodiment humanoid video generation. It offers an interesting perspective on cross-embodiment adaptation, supported by a comprehensive framework and promising experimental results. However, the ablations do not clearly distinguish the benefits of motion–embodiment decoupling from those of embodiment-specific LoRA parameterization. The costs of adapting to new embodiments and the comparability of evaluation settings also require clarification.

Critical strengths
  1. Transferring ideas from image stylization to cross-embodiment tasks. The paper extends LoRA-based disentanglement methods from image style transfer to humanoid video generation, separating shared motion transfer from embodiment-specific appearance. This provides an interesting approach to embodiment heterogeneity and adaptation to new robots.
  2. A comprehensive framework. The work integrates motion transfer, embodiment-specific adaptation, synthetic data construction, and streaming generation into a coherent pipeline. This breadth, supported by synthetic and real-world evaluations, strengthens its value as a practical framework for cross-embodiment video generation.
Critical weaknesses

None

Questions for authors
  1. What evidence distinguishes the benefits of motion–embodiment decoupling from those of embodiment-specific LoRA parameterization?

  2. What data preparation, computational resources, and training time are required to adapt the model to a new embodiment?

  3. How are target-embodiment adaptation data separated from the test set and made available to each baseline?

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

OmniHumanoid transfers motions between humanoid embodiments through a shared source-video motion pathway and embodiment-specific video LoRAs on the denoising pathway. An asymmetric attention mask lets generation read the conditioning stream while preventing the reverse information flow. Embodiment LoRAs are first trained on unpaired videos; shared motion transfer is then learned from aligned synthetic pairs, with target LoRAs frozen. A new embodiment is introduced through its own unpaired adaptation videos. The authors construct 7,200 paired training samples and evaluate 50 held-out synthetic videos and 50 real-world videos. A causal four-step student is distilled from the 50-step teacher. The teacher reaches 25.47 dB PSNR on the synthetic benchmark, while the full streaming student reaches 23.34 dB at 4.96 FPS. Evaluation includes automated video judgments and an appendix user-preference study.

Critical strengths

A key strength is the modular approach to adapting video generation across robot embodiments. Embodiment-specific LoRAs on the denoising branch, combined with asymmetric attention, provide a clear mechanism for separating target-specific appearance and morphology from shared motion processing. This allows a new robot to be introduced using unpaired target videos while keeping the shared motion-transfer model fixed, addressing the practical cost of collecting motion-aligned demonstrations for every embodiment. A second strength is the integration of streaming generation: causal distillation reduces sampling from 50 to four denoising steps and achieves approximately 5 FPS at 720p on an H200 GPU. Together, these contributions address both adaptation cost and generation throughput, making the framework a useful step toward scalable production of robot-centric demonstration videos.

Critical weaknesses

The central paired-free adaptation claim is difficult to assess because the adaptation protocol is insufficiently specified. Although the method adapts an unseen robot using unpaired videos, the paper does not report the exact Unitree-G1 adaptation-set size, its separation from the 50 test videos in terms of motions and scenes, or whether trainable baselines receive equivalent target-embodiment adaptation data. These details are essential to establish generalization and a fair comparison. A second weakness concerns the validation of motion fidelity and embodiment preservation. Whole-frame reconstruction metrics do not isolate robot articulation or contact accuracy, while the reference-free evaluation relies on a single VLM. Appendix B provides encouraging human preferences, but omits participant and clip counts, presentation and blinding procedures, and uncertainty estimates, making the strength of this evidence difficult to judge. These limitations materially affect my assessment because adaptation to new embodiments and faithful motion transfer are the paper’s central contributions, yet the current evaluation does not establish their robustness with sufficient clarity.

Questions for authors
  1. Unseen-embodiment adaptation and comparison fairness. How many Unitree-G1 videos were used to train its LoRA, and how were these separated from the 50 test videos with respect to motions and scenes? Did the trainable baselines receive equivalent unpaired target-embodiment data and adaptation opportunities? Clarifying the shared-training, adaptation, and testing splits would materially strengthen confidence in the central generalization claim.

  2. Human evaluation. For the user study in Appendix B, please specify the participant and clip counts, the clip-selection procedure, and how method identities and presentation order were controlled. Were the evaluated outputs generated by the teacher or by the streaming student, and what uncertainty surrounds the reported preference rates? These details would help establish the reliability and scope of the perceptual advantage.

If these points are addressed properly, I will raise my score.

1 comment

Thank you for the clarification. Two concerns remain:

  1. Adaptation–test separation. Excluding G1 from shared-model training does not establish that its adaptation data are independent of the test set. Are the 50 G1 adaptation videos strictly disjoint from the 50 test videos? Please clarify any overlap in underlying motion sequences and scenes, and specify how the split was constructed.

  2. Human evaluation of the streaming model. Evaluating the teacher against offline baselines is reasonable, but the reported preferences do not establish the perceptual quality of the streaming student, which may degrade after distillation. Please provide a corresponding student evaluation or explicitly restrict the human-preference conclusions to the teacher.

Reviewer_2F9SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OmniHumanoid translates source human/robot videos into videos of a target humanoid embodiment. The method separates a shared motion-conditioning pathway from embodiment-specific denoising-branch LoRAs, with asymmetric attention intended to prevent target-appearance information from contaminating the shared pathway. Embodiment LoRAs are first trained on unpaired videos, then frozen while the shared transfer component learns from aligned cross-embodiment pairs. A new embodiment is introduced by training only its LoRA on unpaired videos. The paper also distills a 50-step bidirectional teacher into a four-step causal streaming student. Validation uses a Unity-rendered dataset with 7,200 paired training examples, 50 held-out synthetic test videos, 50 real-world source videos, VLM-based evaluation, a preference study, and system-level ablations. The reported streaming throughput is about 5 FPS at 720p on one H200, with a quality reduction relative to the teacher.

Critical strengths
  1. A useful modular formulation for adding robot embodiments without new paired demonstrations. Restricting target-specific LoRAs to the denoising branch and blocking reverse information flow is a sensible way to reduce interference between appearance adaptation and shared motion transfer.
Critical weaknesses
  1. The practical value of translating human videos into humanoid robot videos is not sufficiently clear to me. For embodied AI, physical consistency and plausibility are generally more important than photorealistic appearance. This raises concerns about whether the reported evaluation metrics adequately capture the properties relevant to embodied applications. Moreover, the paper does not demonstrate any downstream embodied AI applications.

  2. I was unable to play the videos embedded in the paper, which limits my ability to properly assess the quality of the generated results.

Questions for authors

See weaknesses

Comments on the paper

1 comment

Dear reviewers,

Thanks for serving as a reviewer. As the discussion period comes to a close and the authors have submitted their rebuttals, I kindly ask you to take a moment to review them and provide any final comments.

If you have already updated your comments, please disregard this message.

Thank you once again for your dedication to the OpenReview process.

Best,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.