acceptodds
Under review as a conference paper at ICLR 2027

MoGeFlow: Flowing Through Motion Codebook Geometry for Text-to-Motion Generation

Abstract

Vector-quantized tokenizers have become the dominant interface for text-to- motion generation, and the priors built on them treat motion codes as unordered categorical labels. We show that this discards structure the tokenizer already learned. A motion code embedding is a decoder-bound movement prototype, and learned motion codebooks carry measurable, decoder-causal geometry: code distances track the distances between the movements those codes decode to, the alignment collapses under shuffled controls, and displacement in code space pre- dicts how far the frozen decoder moves. Geometry of this kind calls for a dif- ferent generative object—a continuous trajectory through the code space rather than a sequence of decisions over a vocabulary. We introduce MoGeFlow, a text- conditioned flow over part-structured motion-code frames. Its tokenizer quantizes each body-part group separately, so every decoder input is a point on a combina- torial lattice of decoder-bound states, and the flow regresses onto that lattice. Be- cause the decoder responds smoothly to code-space displacement, generated states are decoded as they are, and the detail the flow places between codes reaches the decoder instead of being rounded away. Quantization therefore serves generation entirely at training time, shaping the space rather than gating the samples: under matched conditions the lattice-shaped target beats both the unquantized latents it is built from and a matched variational latent space. MoGeFlow sets a new state of the art in text–motion alignment on HumanML3D, leads Top-3 retrieval and Mul- tiModal Distance on KIT-ML, and on the large-scale MotionMillion benchmark surpasses a 7B-parameter model in retrieval with an order of magnitude fewer parameters. Quantization shapes the space; the flow generates through it.

Primary area:
generative models
Keywords:
motion generation

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_uvStOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the unordered problem in the VQ method, which is widely used in the T2M field. It utilized a part-aware methodology and conducted MotionMilltion benchmarks and evaluations alongside HumanML3D and KIT-ML, which are commonly used in this field.

Critical strengths
  1. The problem is intuitive.
  • It is intuitively reasonable to think that using the commonly used cross-entropy when training with VQ is a problem, as it applies the same penalty regardless of whether the motion is close or far.
Critical weaknesses
  1. Lack of Prior Research
  • There is a lack of comparative analysis and comparison with existing part-aware methodologies such as ParCo (ECCV 2024), LGTM (SIGGRAPH 2024), BiPO (WACV 2026), and ParTY (CVPR 2026) when using the Part-aware Tokenizer.
  1. Experimental Results Fail to Support the Claims
  • To my knowledge, FID is the most important metric for comparing the quality of generated motion in this field. R-Precision and MultiModal Dist serve to assist in performance evaluation; high scores on these metrics alone cannot support the claim that a model performs well. In particular, R-Precision and MultiModal Dist perform better than Real Motion, making it impossible to make a rational judgment based solely on these metrics. Furthermore, the fact that MultiModality—which demonstrates diversity—shows the lowest performance among the comparative papers is a significant issue, suggesting that the generative model is not fulfilling its role and may be overfitted.
  1. Exaggeration of Efficiency Claims
  • While the model claims efficiency through comparison with the MotionMillion-7B model, there is no comparison with Light-T2M, an existing model in this field that also claims efficiency. In particular, performance is poor in FID, a metric more important than Light-T2M, and an analysis of this is required.
  1. Absence of User Study
  • User study is essential for assessing the quality of human motion, but this paper lacks such an evaluation.
Questions for authors
  1. Please identify the differences between this model and the part-aware tokenizer methodologies used in existing part-aware methodologies, and if possible, conduct comparative experiments and analyses.

  2. Since R-Precision and MultiModal Dist show higher performance than Real Motion, please add an analysis regarding whether the claimed model's performance is superior.

  3. The MultiModality is significantly low; please add an analysis regarding this.

  4. Please measure and compare FLOPs, model parameters, and latency with existing models in the Text-to-Motion field.

  5. There have been many recent assessments that the feature extractor is biased. Please conduct evaluation experiments using the evaluation tool used in MARDM and TMR.

  6. Please conduct user study with at least 20 users.

Reviewer_GmXNOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MoGeFlow is a text-to-motion flow-matching model. It adopts a part-structured VQ tokenizer and trains a DiT to generate motion embeddings from Gaussian noise. At inference, generated continuous motion latent are passed directly to the vq motion decoder without quantization. Its main empirical claim is that quantization provides useful training targets without requiring discrete outputs at inference. Validation includes HumanML3D, KIT-ML, and MotionMillion, and ablations on representation, training targets, model capacity, and decoding mode.

Critical strengths
  1. The paper is easy to reproduce.
  2. MoGeFlow has significant better text–motion alignment.
Critical weaknesses
  1. Limited novelty, with insufficient evidence to establish a broader empirical principle. The main claim of MoGeFlow is that quantization can provide useful training targets without requiring discrete outputs at inference. However, this design principle is not entirely new.

    • LDM [1] already provides a force_not_quantize=True decoding path and explicitly compares quantized and non-quantized decoding of the same generated sample in its official implementation (Code). Also, Kandinsky[2] also evaluates inference with and without latent quantization in Table 3. Although MoGeFlow differs from LDM in using post-quantization embeddings as its training targets, the novelty of the overall design therefore appears limited.
    • More importantly, the empirical finding that training on quantized targets while decoding continuous outputs improves generation quality is not sufficiently established beyond the specific configuration studied. This introduces a potential train–inference mismatch: training targets lie on the codebook lattice, whereas generated terminal states are generally off-lattice. While the reported ablations support this design choice on HumanML3D, it remains unclear how robust the benefit is to changes in the representation. For example, evaluating different codebook sizes on different dataset (especially on MotionMillion scale dataset) would help determine whether the improvement reflects a general property of the proposed approach or is specific to the chosen tokenizer configuration.
  2. The substantial improvement on MotionMillion requires more convincing evaluation and interpretation. Table 2(a) reports a large improvement in retrieval performance over MotionMillion-7B: R@1 increases from 0.79 to 0.93, while FID deteriorates from 10.3 to 12.1. These results are not inherently contradictory, but they leave open whether MoGeFlow achieves better semantic alignment at the expense of motion distribution fidelity.

    • The authors should report the TMR retrieval performance on ground-truth motions under the same evaluation protocol. This reference is important for assessing how close the reported retrieval scores are to the evaluator's attainable performance and how much of the improvement reflects the limitations of the metric.
    • The comparison also raises a potential issue of evaluation fairness: MoGeFlow conditions on ground-truth motion lengths, whereas autoregressive baselines such as MotionMillion may generate sequences with predicted lengths.
    • Retrieval metrics alone are insufficient to establish improved generation quality. TMR-based retrieval scores can be sensitive to evaluation and retrieval settings, including candidate-set size.
    • MoGeFlow's advantage over MotionMillion-7B is not solely attributable to its smaller model size. RMG [3] also reports improved text-to-motion retrieval performance, with a more consistent trade-off between FID and R@K. In comparison, MoGeFlow's substantial R@K gains alongside degraded FID raise questions about whether its retrieval improvements translate into better overall generation quality.

Reference: [1] Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022. [2] Razzhigaev et al. Kandinsky: An Improved Text-to-Image Synthesis with Image Prior and Latent Diffusion. EMNLP 2023 [3] Miao et al. "Riemannian Motion Generation: A Unified Framework for Human Motion Representation and Generation via Riemannian Flow Matching." arXiv preprint arXiv:2603.15016 (2026).

Questions for authors

VAE baseline: For the matched PartVAE comparison in Table 3(b), could the authors report the tokenizer's reconstruction quality with T2M metrics? This would help clarify whether the performance gap reflects the target-space representation itself or differences in reconstruction quality.

Codebook size: Could the authors provide an ablation on different codebook sizes to evaluate whether the benefits of quantized-target training are robust to this design choice?

MotionMillion ablation: Could the authors provide ablations on MotionMillion to verify whether the observed gains persist on this benchmark, particularly given the substantial improvement in retrieval performance but worse FID compared with MotionMillion-7B?

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

This paper investigates the geometry of learned motion codebooks and shows that distances between code embeddings correlate with differences in the corresponding decoded motions. The authors propose MoGeFlow, which trains a text-conditioned flow model to generate continuous motion embeddings using quantized codebook representations as training targets. During inference, the generated embeddings are directly passed to a frozen motion decoder without re-quantization. Experiments on multiple text-to-motion generation benchmarks demonstrate competitive performance, particularly in text-motion alignment.

Critical strengths
  1. It is an interesting idea to explore the correlation between continuous kinematic space and discrete codebook space and find that nearby code embeddings tend to encode similar motions.

  2. The motivation is well supported by correlation analysis, shuffled controls, and code replacement experiments.

  3. Comprehensive experiments are provided. MoGeFlow achieves state-of-the-art text-motion alignment on HumanML3D and competitive results on KIT-ML and MotionMillion.

Critical weaknesses
  1. Although the analysis of codebook geometry is interesting, the actual generation framework does not introduce an explicit geometry-aware mechanism. The codebook geometry is primarily used to motivate the choice of training targets rather than being directly incorporated into the model architecture or optimization objective.

  2. While the paper demonstrates a strong global correlation between codebook embedding distances and decoded motion distances, this does not guarantee local consistency. If some nearby code embeddings represent substantially different motions, could this negatively affect motion generation? Does the proposed method consider or address this potential issue?

  3. The proposed method mainly combines an existing part-structured VQ tokenizer with standard conditional flow matching. The paper would benefit from a clearer explanation of its technical novelty over existing latent diffusion and flow-based motion generation methods.

  4. The paper uses a part-structured VQ tokenizer. Did the authors try any non-part-structured tokenizer? It would be helpful to isolate the contribution of part structures and analyze the generalization to non-part structures.

  5. The paper uses ground-truth motion lengths during inference. It would be helpful to explicitly indicate whether each compared method uses ground-truth motion lengths in Table 1 to ensure a fair and transparent comparison.

Questions for authors
  1. The MultiModality score of MoGeFlow is relatively low compared to some existing methods in Table 1. Could the authors provide an explanation for this?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.