PoseRieFlow: Riemannian Flow Matching for Whole-Body Text-to-Pose Generation
Organizations: University of Bonn
Abstract
We present PoseRieFlow, a compact flow-matching framework for text-conditioned generation of whole-body 3D human poses. Prior text-to-pose methods either neglect articulated hands or model rotations in Euclidean space despite their non-Euclidean geometry. PoseRieFlow instead formulates generation directly on the product manifold of SMPL-H joint rotations, learning a text-conditioned vector field that transports samples from an isotropic Gaussian prior to the data distribution along Riemannian geodesics. A part-tokenized Diffusion Transformer aggregates the 52 joints into eight body-part tokens, allocating a dedicated token and nonlinear projector to each hand, and restores per-joint resolution through part-specific decoders, shortening the sequence while improving kinematic fidelity over per-joint tokenization. With only 21.6M trainable parameters, PoseRieFlow supports expressive SMPL-H generation with detailed hand articulation at a fraction of the scale of recent LLM-based generators. On PoseScript, it attains the lowest whole-body kinematic error among all hand-producing methods, with consistent kinematic gains from the Riemannian formulation over its Euclidean counterpart at every practical inference budget, while distributional and retrieval metrics remain comparable between them.
Figures & tables
| R@1(32) (%) | PA-MPJPE | Geo | Geo ∖0 | Div | ||
| Method | T2P | P2T | (mm) | ( ∘ ) | ( ∘ ) | (mm) |
| PoseScript (VAE) | ||||||
| ChatPose | ||||||
| UniPose | ||||||
| PoseLLaVA | ||||||
| Ours (Eucl.) | ||||||
| R@1(32) (%) | PA-MPJPE | Geo | Hands | Div | ||||
| Method | FID | T2P | P2T | (mm) | ( ∘ ) | Geo ( ∘ ) | TipPA (mm) | (mm) |
| PoseScript (VAE) | ||||||||
| Ours (Eucl.) | ||||||||
| Ours (Riem.) | ||||||||
| NFE | 1 | 2 | 5 | 10 | 20 | 50 | 100 | |
|---|---|---|---|---|---|---|---|---|
| PA (mm) | Eucl. | 167.7 | 143.7 | 113.4 | 112.9 | 114.7 | 116.4 | 117.1 |
| Riem. | 213.6 | 156.4 | 111.5 | 109.3 | 110.4 | 111.6 | 112.1 | |
| Geo (deg) | Eucl. | 14.85 | 12.38 | 10.69 | 10.57 | 10.61 | 10.65 | 10.69 |
| Riem. | 18.72 | 13.04 | 10.52 | 10.30 | 10.28 | 10.31 | 10.33 |
| Variant | FID | R@1 T2P | R@1 P2T | PA-MPJPE | Geo | |
|---|---|---|---|---|---|---|
| (%) | (%) | (mm) | (deg) | |||
| Ours (Eucl.) | 1.0 | 0.947 | 24.01 | 19.41 | 123.8 | 10.61 |
| 2.0 | 0.903 | 27.54 | 22.64 | 116.4 | 10.65 | |
| 3.0 | 0.887 | 28.47 | 23.72 | 117.9 | 10.92 | |
| Ours (Riem.) | 1.0 | 0.960 | 24.46 | 19.22 | 120.7 | 10.39 |
| 2.0 | 0.942 | 27.42 | 22.29 | 111.6 | 10.31 |
| Token | Train Data | FID | R@1 T2P | R@1 P2T | PA-MPJPE | Geo |
|---|---|---|---|---|---|---|
| (%) | (%) | (mm) | (deg) | |||
| Joint | PS only | 0.926 | 28.18 | 22.82 | 114.1 | 10.60 |
| Joint | PS+MX++ | 0.929 | 28.11 | 22.65 | 117.9 | 10.85 |
| Part | PS only | 0.942 | 27.42 | 22.29 | 111.6 | 10.31 |
| Part | PS+MX++ | 0.950 | 27.91 | 22.44 | 112.6 | 10.54 |
| Encoder | Cond. | FID | R@1 T2P | R@1 P2T | PA-MPJPE | Geo |
|---|---|---|---|---|---|---|
| (%) | (%) | (mm) | (deg) | |||
| Qwen3-Emb-4B | pooled | 0.978 | 18.05 | 14.22 | 163.0 | 12.09 |
| Flan-T5-Base | pooled | 0.974 | 17.48 | 14.04 | 164.4 | 12.17 |
| Flan-T5-Base | cross | 0.942 | 27.42 | 22.29 | 111.6 | 10.31 |
Text-to-3D Generation
Text-to-3D generation creates 3D objects and scenes from language while balancing semantic alignment, geometry, appearance, and spatial coherence.
See all 69 papers