acceptodds
Withdrawn from ICLR 2027

PrefixMotion: Prefix-Consistent Conditional Motion Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Interactive conditional motion generation requires new observations to extend a motion sequence without revising its history. We introduce PrefixMotion, a framework that adapts pretrained motion generators to prefix-consistent streaming inference. Under compatible conditions and coupled randomness, full-sequence and prefix executions agree on their shared outputs. The framework enforces this agreement across attention, shape aggregation, and motion decoding. A clean-motion parameterization of rectified flow preserves compatibility with pretrained weights and geometric supervision while enabling three-evaluation Euler sampling. Instantiated with GENMO, PrefixMotion is evaluated through video-conditioned reconstruction and text-conditioned synthesis. On 120-frame windows from 37 3DPW tracks, four prefix-length comparisons yield maximum tensor deviations below through the global output, given compatible supplied track and camera prefixes. Cached execution is evaluated on 79 complete sequences and accelerates newest-frame updates by 10.6× over window recomputation on the same GPU. PrefixMotion improves all three 3DPW spatial metrics and all three EMDB-2 trajectory metrics over our evaluation of pretrained GENMO. It also achieves lower error on 10 of 13 metrics than OnlineHMR under shared finalized tracks and DROID-SLAM cameras, with model-specific EMDB intrinsics. On HumanML3D, five-trial mean FID improves from 10.169 to 8.110 and R@3 from 0.475 to 0.492. These results show that explicit temporal constraints can support efficient streaming while preserving conditional generation and improving selected reconstruction metrics.

Primary area:
generative models
Keywords:
human motion generation, prefix consistency, streaming inference, rectified flow, conditional generative models, human mesh recovery, multimodal generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_U4c7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PrefixMotion, a framework for adapting pretrained conditional motion generators to prefix-consistent streaming inference. The central objective is to ensure that previously generated motion remains unchanged when new observations or additional frames become available. Built upon the pretrained GENMO model, PrefixMotion introduces causal temporal attention, running-prefix shape aggregation, and causal video-conditioning and decoding operations. The authors further adopt a clean-motion rectified-flow formulation with a three-evaluation Euler sampler and use key-value caching for incremental inference. The proposed implementation achieves a reported 10.6× speedup over repeated window computation while maintaining numerical agreement with full-sequence inference under the stated assumptions.

Critical strengths
  1. Careful treatment of consistency beyond causal attention. The authors correctly recognize that causal attention alone does not ensure end-to-end prefix consistency. Their treatment of shape aggregation, camera-feature alignment, motion decoding, and temporal postprocessing provides a useful system-level perspective.

  2. Detailed numerical verification and efficient implementation. The paper evaluates prefix agreement at multiple intermediate representations and verifies cached streaming over 79 complete sequences. The reported reduction in per-update latency from 603 ms to 57 ms is practically meaningful.

  3. Broad evaluation across multiple conditioning modalities. The extensive supplementary experiments and ablations demonstrate considerable implementation and evaluation effort.

Critical weaknesses
  1. Limited methodological novelty beyond established causal modeling techniques. The main architectural changes rely on existing techniques, including causal attention masks, key-value caching, running temporal statistics, clean-sample prediction, and rectified-flow sampling. These components are already well established in autoregressive transformers, streaming sequence models, and diffusion-based generation. Related approaches such as Diffusion Forcing (NeurIPS 2024) and Vid2World (ICLR 2026) have also explored per-frame noise levels and causal adaptations of pretrained generative models.

  2. The prefix-consistency guarantee depends on restrictive assumptions about external conditioning. The practical online systems frequently revise tracking associations and camera estimates as additional observations become available. The evaluation primarily uses finalized upstream trajectories and tracks rather than a fully incremental perception pipeline. Consequently, I concern the formal guarantee applies to a constrained setting and does not establish end-to-end consistency in realistic online motion reconstruction.

  3. The effect of causal adaptation on motion-generation quality is not sufficiently isolated. PrefixMotion modifies multiple components simultaneously, including temporal attention, training objectives, flow parameterization, sampling, and postprocessing. The reported quality improvements over pretrained GENMO therefore cannot be attributed specifically to prefix-consistent modeling. It also performs worse than pretrained GENMO on all four EMDB-1 metrics, 3DPW acceleration error, and EMDB-2 jitter and foot sliding.

Questions for authors

Q1. Novelty relative to existing causal generative models. Could the authors clarify which components of PrefixMotion constitute fundamentally new methodological contributions, rather than applications of established causal attention, temporal caching, running statistics, and clean-motion prediction?

Q2. Isolation of causal modeling effects. Could the authors provide a controlled comparison between causal and bidirectional rectified-flow models trained with identical data, objectives, model capacity, and sampling procedures? This would help establish whether the reported reconstruction improvements are attributable to causal temporal modeling rather than changes in the training objective, sampler, or postprocessing.

Q3. Reconstruction quality and temporal consistency trade-offs. PrefixMotion improves several spatial metrics but worsens acceleration, jitter, and foot-sliding metrics compared with pretrained GENMO. Could the authors explain whether these degradations arise primarily from causal attention, the NFE-3 Euler sampler, or the modified motion-decoding procedure?

Q4. Practical streaming efficiency. The reported speedup is measured against repeated 120-frame window recomputation. Could the authors provide a comparison with an existing streaming motion estimator or a matched causal implementation that uses key-value caching? It would also be useful to report whether the stated inference latency remains practical when person tracking and camera estimation are included in the complete online pipeline.

1 comment

Thank you for your thorough review and constructive feedback. We appreciate your recognition of our careful treatment of consistency beyond causal attention, detailed numerical verification and efficient implementation, and broad evaluation across multiple conditioning modalities.

Your comments have highlighted several important aspects that warrant further clarification and validation. After considering the scope of the necessary revisions and additional validation, we have decided to withdraw this submission. The amount of further work required is substantial, and our current resources are insufficient to address these issues adequately at this stage.

We sincerely appreciate your time and valuable feedback.

Reviewer_i9QqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue of prefix consistency in conditional human motion generation—ensuring that previously generated motion remains unchanged as new observations are incorporated. The authors propose "PrefixMotion," which builds upon the pre-trained GENMO model by employing causal attention and a running-style human shape aggregation mechanism. Additionally, the post-processing steps for camera and motion are adapted to be causal, thereby achieving prefix-consistent motion generation. The authors also integrate a clean-motion predictor into a Rectified Flow framework, utilizing three-step Euler sampling and a KV cache to minimize redundant computations during streaming inference.

The study demonstrates that, given compatible input prefixes and shared random noise, the proposed method maintains prefix consistency. Evaluations conducted on the 3DPW, EMDB, and HumanML3D datasets show improvements in both partial reconstruction and generation metrics. Furthermore, the authors validated the caching mechanism across 79 complete test sequences, achieving a 10.6x speedup in update operations compared to window-based recomputation.

Critical strengths
  1. The systematic handling of prefix consistency is quite comprehensive. The paper points out that modifying the attention mechanism alone is insufficient to guarantee prefix consistency in the final output, as steps such as human shape aggregation, camera alignment, and post-processing can still introduce information from future frames. The authors have adapted these operations accordingly—a claim supported by both theoretical analysis and component-wise experiments—thereby establishing prefix consistency as a property that can be explicitly verified throughout the entire generation pipeline. This represents the paper’s most valuable technical contribution.

  2. The acceleration of streaming inference holds practical value. Using an L40S GPU with a 120-frame context, the cache-based execution reduces the latency of single-frame updates from 603 ms to 57 ms—a 10.6x speedup compared to window-based recomputation. The authors also conducted numerical consistency checks across 79 complete test sequences, noting a single instance of a numerical anomaly during the final decoding stage. These results provide compelling evidence for the computational efficiency and engineering feasibility of the proposed method.

Critical weaknesses
  1. The methodological innovation and the specific contributions of key design elements remain insufficiently clear. The paper primarily integrates existing techniques—such as causal attention, running parameter aggregation, Rectified Flow, and KV Cache—with its overall innovation lying largely in the systematic adaptation of a pre-trained model. While these design choices collectively achieve prefix consistency, the main text lacks a strictly comparable bidirectional Flow Matching baseline; consequently, it is impossible to isolate the impact that removing future context has on generation quality. As a result, the actual contribution of the core methodological modifications remains ambiguous.

  2. There is a clear trade-off regarding generation quality, yet the associated costs have not been fully articulated. Although PrefixMotion improves certain spatial and trajectory-based metrics, it underperforms compared to GENMO—specifically regarding acceleration error on the 3DPW dataset, as well as across all four metrics on the EMDB-1 dataset. On EMDB-2, the Jitter metric increases from 19.826 to 46.279, and Foot Sliding rises from 4.584 to 11.425. While the authors analyze some of these performance trade-offs through a comparison of different samplers, it remains difficult to disentangle the specific effects on quality stemming from the causal formulation, the training objective, and the sampling strategy, respectively.

Questions for authors

Q1. Could the authors provide a bidirectional Flow Matching baseline that matches the proposed method in terms of training objectives, data, training budget, and sampler? Without such a comparison, it is difficult to assess the impact of removing future context on generation quality, or the practical cost of enforcing causality.

Q2. Some dynamic quality metrics in Table 1 show significant degradation, and Figure 3 already demonstrates that the choice of sampler affects spatial accuracy and dynamic stability. Could the authors elaborate on the primary advantages of the proposed method—given a fixed inference computational budget—compared to other configurations, and explain how the trade-off between quality and efficiency should be interpreted?

Q3. The current prefix-consistency theorem requires external conditions to have compatible prefixes and relies on shared random noise. In practical online systems, scenarios such as camera trajectory adjustments or changes in text conditioning may arise. Could the authors clarify the scope of this guarantee, and specify under what circumstances it would still be necessary to recompute historical outputs?

1 comment

Thank you for your careful and detailed review. We appreciate your recognition that the systematic handling of prefix consistency is quite comprehensive and that the acceleration of streaming inference holds practical value.

Your comments have highlighted several important aspects that warrant further clarification and validation. After considering the scope of the necessary revisions and additional validation, we have decided to withdraw this submission. The amount of further work required is substantial, and our current resources are insufficient to address these issues adequately at this stage.

We sincerely appreciate your time and valuable feedback.

Reviewer_bA7yOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a generic prefix-consistent architecture for existing pretrained motion generators. Presented approach uses rectified flow architecture that remains compatible with existing pretrained weights.
Authors benchmark their approach by showing results of motion generation on video-contioned and text-conditioned motion generation and compare results with one baseline i.e. GENMO.

Critical strengths
  • I believe this paper addresses an important problem i.e. making pretrained priors into streaming prefix-consistent variants and the approach presents a nice formulation for the task too.
  • The problem formulation described in the paper is quite clear.
  • The approach has potential to be general (albeit it is shown with only one method).
Critical weaknesses
  • The biggest weakness of the paper feels to be writing or more specifically: lack of motivation at certain places. For example in abstract, introduction the method and approach are described directly without explaining implication of the problem (ln 35-40) and how the alternatives fare for this specific task (Ln. 41-46) . The method section also needs a rewrite with explaining motivation at each stage.

  • Lack of comparison with baselines (specifically offline baselines) : The paper only compares with GENMO. While I understand that the method uses GENMO, but there are other baselines for the same tasks (both for video conditioned generation and text conditioned generation). For example, text conditioned motion generation: MDM, T2M-GPT or how does the method compare with offline mesh recovery methods like SAM-3D body.

  • The ablation section is not properly present in the main paper. For example, how do authors justify the design choices made in the paper.

  • The discussion section should properly explain the nuances of designing an online method compared to an offline method.

  • one limitation which authors also admit is that the paper only shows results with GENMO, this can be improved.

Questions for authors
  • How does the method compare to vanilla seed-conditioned motion generation i.e fixing prefix during generation (without contributions made by paper)?

  • Can author provide latency comparison with baselines for example OnlineHMR?

  • The video results show lot of foot floating and foot sliding compared to GENMO, can authors explain why is that the case? and can video results be extended to compare with OnlineHMR?

1 comment

Thank you for your thoughtful review and constructive suggestions. We appreciate your recognition that our paper addresses an important problem, presents a nice formulation for the task, and has the potential to be general.

Your feedback has highlighted several aspects that could be further improved. After considering the scope of the necessary revisions and additional validation, we have decided to withdraw this submission. The amount of further work required is substantial, and our current resources are insufficient to address these issues adequately at this stage.

We sincerely appreciate your time and valuable feedback.

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

Related papers

Discussion (0)

Sign in to comment.