Research map / Image generation / Fast diffusion sampling
Riemannian MeanFlow
Dongyeop Woo, Marta Skreta, Seonghyun Park, Kirill Neklyudov, Sungsoo Ahn
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| use x1-prediction instead of v-prediction for the flow map (ablation) | Semigroup RMF · spherical helix, D=512 and 2048 | one-step sample quality | Improved |
| use xt-prediction instead of x1-prediction (ablation) | Semigroup RMF · spherical helix, D=512 and 2048 | one-step sample quality | Hurt |
| use semigroup objective instead of Eulerian or Lagrangian (ablation) | RMF · spherical helix, D=512, v-pred | sample quality | Improved |
| add adaptive loss weighting (p=0.5) (ablation) | Eulerian RMF · spherical helix, D=512 | sample quality | Improved |
| use low-frequency time embeddings (omega 0.02) instead of high-frequency (30) (ablation) | RMF differential objectives · training stability | stability | Improved |
| place stop-gradient on the whole target so no higher-order derivatives are needed | RMF vs GFM · training memory | memory usage: nearly 2x lower | Improved |
| one-step RMF instead of 100-step Fisher FM | promoter DNA design · FANTOM5 promoter, MSE | MSE: 0.030 -> 0.030 | No clear effect |
| use RMF instead of FrameDiff and FrameFlow at 5 NFE | protein backbone generation · SCOPe, 5 NFE | designability: 0.04 -> 0.82 vs FrameFlow | Improved |
| optimize for designability | RMF · SCOPe | novelty | Hurt |
| evaluate reward on x1 look-ahead instead of current state xt (ablation) | reward-guided RMF · promoter DNA, 5 NFE | MSE: 0.017 -> 0.013 | Improved |
| evaluate reward on x1 look-ahead instead of current state xt (ablation) | reward-guided RMF · protein alpha-helix, 10 NFE | mean alpha-helix fraction: 0.29 -> 0.45 | Improved |
| add intermediate inference noise (eta 0.25-0.45) (ablation) | RMF protein sampling · SCOPe | scRMSD and diversity | Improved |
Closest papers by meaning
- Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation (NeurIPS 2026 · similarity 0.87)
- Fisher Flow Matching for Generative Modeling over Discrete Data (NeurIPS 2024 · similarity 0.87)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.87)
- Learning Distributions on Manifolds with Free-Form Flows (NeurIPS 2024 · similarity 0.87)
- Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design (ICML 2024 · similarity 0.86)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.86)
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.
