acceptodds
Under review as a conference paper at ICLR 2027

Riemannian Motion Generation: A Unified Framework for Human Motion Representation and Generation via Riemannian Flow Matching

Abstract

Human motion generation is often learned in Euclidean spaces, although valid motions follow structured non-Euclidean geometry. We present Riemannian Motion Generation (RMG), a unified framework that represents motion on a product manifold and learns dynamics via Riemannian flow matching. RMG factorizes motion into several manifold factors, yielding a scale-free representation with intrinsic normalization, and uses geodesic interpolation, tangent-space supervision, and manifold-preserving ODE integration for training and sampling. On HumanML3D, RMG achieves state-of-the-art FID in the HumanML3D format (0.043) and ranks first on all reported metrics under the MotionStreamer format. On MotionMillion, it also surpasses strong baselines (FID 5.6, R@1 0.86). Ablations show that the compact (translation + rotations) representation is the most stable and effective, highlighting geometry-aware modeling as a practical and scalable route to high-fidelity motion generation.

Primary area:
generative models
Keywords:
Human Motion Generation, Text-to-Motion Synthesis, Motion Representation Learning, Generative Modeling, Riemannian Flow Matching

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

This paper investigates the issue of geometric modeling in the representation and generation of human motion. The authors point out that existing methods typically employ representations in Euclidean space that contain redundant information, failing to fully exploit the intrinsic geometric structure of human movement. To address this, the paper proposes "Riemannian Motion Generation" (RMG), which decomposes motion into global translation and joint rotations, and generates motion via Riemannian flow matching on a product manifold. The framework integrates geodesic interpolation, supervision of velocities in the tangent space, and ODE integration that preserves manifold constraints, resulting in a unified training and sampling process.

On the HumanML3D dataset, this method achieves an FID of 0.043; on MotionMillion, it attains an FID of 5.6 and an R@1 of 0.86. Ablation studies further demonstrate that using a compact representation—consisting solely of translation and rotation—outperforms configurations that incorporate additional positional or temporal difference terms, both in terms of generation quality and guidance stability.

Critical strengths
  1. Unified and compact geometric modeling of motion. The paper organizes global translations and joint rotations into a product manifold and integrates this with Riemannian Flow Matching, ensuring a consistent geometric foundation for both motion representation and the generation process. This design reduces redundancy in the representation; ablation studies further demonstrate that this compact representation of translation and rotation leads to superior generation quality and improved guidance stability.

  2. Strong experimental performance with adaptability to varying data scales. On the HumanML3D dataset, RMG achieves an FID of 0.043—slightly outperforming the 0.045 reported for MoMask—while maintaining competitive results in terms of text-motion alignment and diversity. On the MotionMillion dataset, RMG attains an FID of 5.6, significantly surpassing the 10.3 achieved by the strongest baseline. These results underscore the framework's effectiveness across different data scales.

Critical weaknesses
  1. Lack of direct validation for the core geometric contribution. While the paper posits that modeling on a Riemannian manifold enhances human motion generation, the existing ablation studies primarily compare different representation components. There is no direct comparison against standard Euclidean Flow Matching—using the same translation/rotation representations, network architecture, and training budget—to isolate whether performance gains stem from the compact representation or the Riemannian generation process itself. This represents a significant evidentiary gap regarding the paper’s central claim and complicates the assessment of the method's true innovation.

  2. Insufficient validation of the practical benefits of geometric modeling. The paper employs unit quaternions to represent rotations; however, it remains unclear how sign-convention consistency is maintained during prior sampling, geodesic interpolation, and inference. Furthermore, while manifold constraints ensure the validity of rotation representations, they do not inherently guarantee that the generated human motions adhere to physical constraints, such as joint range-of-motion limits or contact constraints. The current experiments lack direct metrics—such as foot-sliding, violations of joint constraints, or temporal continuity—making it difficult to determine whether the geometric modeling approach actually improves these specific aspects of motion quality.

Questions for authors

Q1. Regarding the contribution of the core methodology: Could the authors compare Euclidean Flow Matching with Riemannian Flow Matching while keeping the representation of translation and rotation, the network architecture, and the training budget constant? This experiment would be crucial for assessing the independent contribution of the geometric modeling approach.

Q2. Regarding the handling of quaternions: How do the authors consistently account for the equivalence between qq and −q-q across training data, prior sampling, geodesic interpolation, and ODE integration? Specifically, are there any issues with discontinuities or numerical instability, particularly for rotations approaching 180∘180^\circ?

Q3. Regarding ablation studies and motion quality: Since different motion representations require distinct conversion pipelines to be evaluated within a unified format, how do the authors ensure that errors introduced during these conversions do not confound the ablation results? Furthermore, are there metrics—such as foot sliding or joint range-of-motion—that could directly substantiate the claim that the geometric modeling approach improves the plausibility of the generated motion?

Reviewer_5u7BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper generates human motion with flow matching on root translations and joint quaternions. It uses a wrapped Gaussian prior and geodesic interpolation, predicts tangent velocities, and integrates them with exponential maps. The method is evaluated on HumanML3D and MotionMillion, with ablations of the motion representation.

Critical strengths
  1. The main contribution is a complete pipeline that combines a compact motion representation with Riemannian flow matching. Interpolation, velocity supervision, and integration all use the quaternion geometry. Exponential-map updates preserve unit quaternions. The representation ablations favor using translation and rotations alone.
  2. The MotionMillion results are promising. At guidance 2.0, FID is 5.6 and R@1 is 0.81, compared with 10.3 and 0.79 for MotionMillion-7B. Guidance 3.0 raises R@1 to 0.86, with FID 7.8. HumanML3D results include 20 runs and 95% confidence intervals.
Critical weaknesses
  1. There is no Euclidean flow-matching comparison with the same compact representation, backbone, text encoder, and training budget. Figure 3 changes the representation within RMG, so it does not isolate the benefit of Riemannian dynamics. The gains could also come from the compact representation or model configuration. On HumanML3D, FID is 0.043±0.002 versus MoMask's 0.045, without a baseline interval to assess the small difference. MotionMillion reports only one run. The full system works well, but the contribution of its main geometric design remains unclear.
  2. Quaternion sign handling needs clarification. Appendix D uses q0 ≥ 0 and computes the spherical angle from the quaternion inner product. Nearby rotations on opposite sides of 180° can have almost opposite quaternion representatives under this convention. The convention alone does not ensure the shortest rotation path. The paper does not explain whether it aligns signs between quaternion pairs or across time, or how it handles nearly antipodal pairs. These details affect the claimed stability of interpolation.
Questions for authors
  1. How does RMG compare with Euclidean flow matching when the representation, backbone, text encoder, training data, and sampling budget are fixed?
  2. Beyond q0 ≥ 0, what sign alignment and numerical safeguards are used for interpolation, especially for motions crossing 180°?
  3. Do the 20 HumanML3D runs vary training seeds or only generation seeds? Which reported checkpoints use MotionMillion pretraining?
  4. Please specify translation normalization, relative factor weighting, and the integration schedule. Also, could you correct the numerical examples on p. 8 that do not match the curves in Figure 3a?
Reviewer_rYSROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper adapts Riemannian flow matching to human motion generation. The proposed method was tested on HumanML3D and MotionMillion.

Critical strengths

I generally like the idea; properly using Riemannian geometry to model poses and motions makes sense. The paper is also well written, and reading it is quite pleasant. I just wish the methodology were correctly described (see the first Critical Weakness).

Critical weaknesses

As it stands, the whole method section describes a pose manifold and pose Riemannian flow matching, which was already introduced in [1] (the paper currently describes [1]). Even Fig.2 technically depicts a pose manifold, with the trajectory being the motion. If I read between the lines, my understanding is that the proposed method is indeed a motion manifold, but the description is wrong. My interpretation of what is actually going on is the following: the manifold on which RFM operates is the L-fold product manifold of the one described, where L is the number of frames, thus (R3×(S3)J)L(\mathbb{R}^3 \times (\mathbb{S}^3)^J)^L. This would consequently cascade across the whole notation and narrative, with the equation at line 294 also being applied across all joints and frames. The figure should either represent a single point on this motion manifold or, alternatively, the authors could stack L manifolds with the trajectory evolving from manifold to manifold. Note that this is my main criterion for such a strong rejection.

Additionally, contributions seem to be rather limited for the venue considering that Euclidean flow matching was already introduced in [2], PoseRFM (method erroneously described by the authors) was introduced in [1], and other properly formulated Riemannian frameworks for motion generation already exist [3]. These works should be at least acknowledged, and a comparison against [3] would also be welcome. Note that depending on the rebuttal and discussion, I am generally willing to increase my score.


[1] Nadar, J., Foti, S., & Birdal, T. (2026). Posed-flow: Versatile and guided flow matching model of human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 21165-21175).

[2] Hu, V. T., Yin, W., Ma, P., Chen, Y., Fernando, B., Asano, Y. M., ... & Snoek, C. G. (2023). Motion flow matching for human motion synthesis and editing. arXiv preprint arXiv:2312.08895.

[3] Yu, Z., Foti, S., Zhang, L., Zhao, A., Keskin, C., Zafeiriou, S., & Birdal, T. (2026). Geometric neural distance fields for learning human motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2232-2242).

Questions for authors
  1. The method leverages Diffusion Transformer (line 340), was the attention between frames? If yes, I think it would be worth discussing it more.

  2. Two of the four listed contributions are essentially related to results. I would claim that experiments and results are an integral part of the scientific process, thus a necessary requirement to substantiate the claims of the paper rather than a contribution. A new metric or new way of evaluating the model would be a novel contribution, but this is not the case.

  3. Just to confirm (I think the answer is yes), but are all methods in Tab.2 trained on the same dataset? Additionally, are they matched in any way (e.g., number of learnable parameters, training time, etc.)? I think this would be very important for a fair comparison, otherwise authors should prove they achieve better results despite using fewer parameters and training for less.

  4. Why are so many metrics missing in Tab.2?

  5. If my interpretation of the actual model is correct (see Critical Weakness 1), and one point corresponds to a full motion sequence, can the model generate longer sequences without retraining the model to generate more frames?

Comments on the paper

1 comment

I appreciate your careful response, which has addressed most of my concerns. I will keep my score unchanged, and I am happy to leave the final decision to the AC / co-reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.