acceptodds
Under review as a conference paper at ICLR 2027

FlowCoMotion: Text-to-Motion Generation via Token–Latent Flow Modeling

Abstract

Text-to-motion generation is driven by learning motion representations for semantic alignment with language. Existing methods rely on either continuous or discrete motion representations. However, continuous representations entangle semantics with dynamics, while discrete representations lose fine-grained motion details. In this context, we propose FlowCoMotion, a novel motion generation framework that unifies both treatments from a modeling perspective. Specifically, FlowCoMotion employs token-latent coupling to capture both semantic content and high-fidelity motion details. In the latent branch, we apply multi-view distillation to regularize the continuous latent space, while in the token branch we use discrete temporal resolution quantization to extract high-level semantic cues. The motion latent is then obtained by combining the representations from the two branches through a token-latent coupling network. Subsequently, a velocity field is predicted based on the textual conditions. An ODE solver integrates this velocity field from a simple prior, thereby guiding the sample to the potential state of the target motion. Extensive experiments show that FlowCoMotion achieves competitive performance on text-to-motion benchmarks, including HumanML3D and SnapMoGen.

Primary area:
generative models
Keywords:
Text-to-Motion, Motion Generation, Generative Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TekoOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowCoMotion, a text-to-motion generation framework that combines an RVQ-based discrete token representation with a continuous VAE latent representation, followed by text-conditioned flow matching. The two branches are fused through a token-latent coupling network, and the resulting representation is used as the target latent for generation. The framework is trained in three stages and evaluated on HumanML3D and SnapMoGen.

Critical strengths

S1. The overall idea is simple and intuitive to understand.

S2. The evaluation is relatively broad, including two datasets, ablation studies, efficiency analysis, and human evaluation.

Critical weaknesses

W1. The overall novelty is incremental and efficiency claims are not fully convincing. The framework mainly combines already established components, including RVQ-style quantization, continuous VAE latents, and flow matching. It also requires a three-stage training procedure, while the efficiency analysis mainly focuses on inference rather than total training cost. Given the relatively modest gains over strong baselines, the added complexity is not fully justified.

W2. The central token-latent coupling mechanism is insufficiently specified. The paper defines zgt=C(zq,zc)z_{gt} = \mathcal{C}(z_q, z_c), but does not clearly explain how C\mathcal{C} fuses the two representations. Since this coupling is one of the main contributions, the lack of architectural detail makes its technical novelty and effectiveness difficult to assess.

Questions for authors
  • What is the exact architecture and fusion operation of the coupling network C\mathcal{C}?

  • Could the authors report the total training time or computational cost for the full three-stage training pipeline?

  • Could the authors provide supplementary videos showing the generated human motions? It would help to see the effectiveness of the method.

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

This paper proposes FlowCoMotion, a text-to-motion generation framework that integrates discrete motion tokens and continuous latent representations through token-latent coupling. It employs multi-view distillation to enhance motion representations and conditional flow matching for motion generation. Experiments on HumanML3D and SnapMoGen demonstrate competitive performance in motion quality and text-motion alignment.

Critical strengths
  1. This paper is well-written, clearly structured, and easy to follow.
  2. It presents extensive experiments that demonstrate the effectiveness of the proposed method.
Critical weaknesses
  1. The novelty of the paper is limited. Specifically, the proposed combination of discrete and continuous representations was already a well-established paradigm in unified generation and understanding last year, e.g., CODA. In the generation domain, DC-Motion has also proposed a similar idea. In addition, flow matching methods have been applied to motion generation since 2025. The Nova and Orbit modules are standard fusion modules in generative modeling and lack novelty.
  2. Compared with StableMoFusion, LaMP, SALAD, and MotionHiFlow, the experimental results show no advantage in Top-1 and FID, so the claimed state-of-the-art performance cannot be verified.
  3. The current experiments can demonstrate the effectiveness of the overall loss, but they lack validation of Multi-view Distillation, which is one of the main contributions of the paper.
  4. The paper lacks sufficient validation of its core fusion strategy. Although the hybrid representation outperforms single representations, comparisons with common strategies such as concatenation, linear fusion, and cross-attention are missing, making it difficult to demonstrate the independent advantage of the Token–Latent Coupling Network.
Questions for authors

See Weaknesses.

Reviewer_9PAYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowCoMotion, which combines token and latent representations for text-to-motion generation. It encodes motion sequences through a token branch and a latent branch, respectively. The two representations are integrated through a token-latent coupling network, followed by a text-conditioned flow matching model for motion generation. Experiments on HumanML3D and SnapMoGen demonstrate competitive performance compared with existing text-to-motion generation methods.

Critical strengths
  1. The method seems technically sound by combing multiple motion generation techniques, such as token-based discrete representation, continuous latent representations and flow matching.

  2. The methos achieves competitive quantitative motion generation results on HumanML3D.

  3. The visualization examples and comparison with SOTA methods show good qualitative generation quality.

Critical weaknesses
  1. The complementarity of the two representations is not sufficiently validated. The paper argues that discrete tokens primarily capture motion semantics, while continuous latents preserve fine-grained dynamics. Although the probing experiments in Appendix Table B provide some analysis for this distinction, they do not establish whether the two branches actually play these respective roles during motion generation. More importantly, the paper does not provide sufficient evidence to support its central motivation that continuous representations entangle semantics with dynamics, while discrete representations lose fine-grained motion details (Lines 13-14). In fact, existing discrete representation-based methods have demonstrated strong performance in generating fine-grained motion details, sometimes outperforming continuous representation-based methods. More direct experiments, such as branch swapping or intervention-based analysis, would help clarify how each representation contributes to the generated motions.

  2. The method is technically incremental without sufficiently novel new techniques. It mainly combines established motion generation techniques, such as token-based discrete representations, continuous latent representations and flow matching. The paper would benefit from a clearer statement of novelty in terms of findings and tehcniques.

  3. The structure of token-latent coupling network is not clear. It seems a simple combination of two representations without sufficient justification on whether and how it can effectively exploit different properties of two representations.

  4. Why does the paper not provide experimental results on the commonly used KIT benchmark?

Questions for authors

Please see the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.