Research map / 3D vision / Face and shape
Distribution-Aligned Diffusion for Human Mesh Recovery
Lin Geng Foo, Jia Gong, Hossein Rahmani, Jun Liu
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| frame mesh recovery as reverse diffusion instead of single-pass or stacked prediction (ablation) | HMDiff · 3DPW | MPVE (mm): 104.0 -> 82.4 (Baseline A), 97.1 -> 82.4 (Baseline B) | Improved |
| add Distribution Alignment Technique to standard diffusion (ablation) | HMDiff · 3DPW | MPVE (mm): 94.9 -> 82.4 | Improved |
| add DAT, reducing diffusion steps needed (ablation) | HMDiff · 3DPW | diffusion steps needed: 200 -> 40 | Improved |
| feed pose heatmap as input feature instead of using DAT (ablation) | standard diffusion · 3DPW | MPVE (mm): 94.9 -> 92.7 (still far from 82.4) | No clear effect |
| align via gradient on estimated h0 instead of directly modifying h_k with L2 loss (ablation) | DAT · 3DPW | MPVE (mm): 91.3 -> 82.4 | Improved |
| use full pose distribution U as prior rather than a single pose u_c (ablation) | DAT · 3DPW | MPVE (mm): 89.9 -> 82.4 | Improved |
| add activation strategy that deactivates DAT once the relative gap is below threshold (ablation) | DAT · 3DPW | MPVE (mm): 85.3 -> 82.4 | Improved |
| start reverse process from Gaussian noise rather than upsampled prior distribution H_U (ablation) | DAT · 3DPW | MPVE (mm): 93.5 -> 82.4 | Improved |
| HMDiff versus FastMETRO inference | HMDiff · Human3.6M | FPS: 23 -> 18 FPS; MPJPE 52.2 -> 49.3 | Hurt |
Closest papers by meaning
- DPMesh: Exploiting Diffusion Prior for Occluded Human Mesh Recovery (CVPR 2024 · similarity 0.90)
- ADHMR: Aligning Diffusion-based Human Mesh Recovery via Direct Preference Optimization (ICML 2025 · similarity 0.90)
- Human-3Diffusion: Realistic Avatar Creation via Explicit 3D Consistent Diffusion Models (NeurIPS 2024 · similarity 0.88)
- DiSRT-In-Bed: Diffusion-Based Sim-to-Real Transfer Framework for In-Bed Human Mesh Recovery (CVPR 2025 · similarity 0.88)
- DiffHuman: Probabilistic Photorealistic 3D Reconstruction of Humans (CVPR 2024 · similarity 0.88)
- Score-Guided Diffusion for 3D Human Recovery (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
