Research map / Image generation / Fast diffusion sampling
RMFlow: Refined Mean Flow by a Noise-Injection Step for Multimodal Generation
Yuhao Huang, Shih-Hsin Wang, Andrea L. Bertozzi, Bao Wang
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 |
|---|---|---|---|
| add noise-injection refinement and likelihood loss to 1-NFE MeanFlow | RMFlow · COCO FID-30K | FID: 27.31 -> 18.91 | Improved |
| add noise-injection refinement and likelihood loss to 1-NFE MeanFlow | RMFlow · COCO2017 CLIP score | CLIP score: 0.273 -> 0.291 | Improved |
| add noise-injection refinement and likelihood loss | RMFlow w/ contexts · QM9 | Molecule stability (%): 84.3 -> 93.2 | Improved |
| add context encoding (ablation) | RMFlow · QM9 | Molecule stability (%): 90.1 -> 93.2 | Improved |
| add policy-gradient physical feedback (RLPF) | RMFlow w/ contexts · QM9 | Molecule stability (%): 93.2 -> 93.5 | No clear effect |
| add noise-injection refinement to 1-NFE MeanFlow | RMFlow · 1D Gaussian mixture | KL divergence: 0.8074 -> 0.2332 | Improved |
| 1-NFE RMFlow vs 32-NFE MeanFlow | RMFlow · 1D Gaussian mixture | KL divergence: 0.1017 vs 0.2332 | Hurt |
| add RMFlow refinement on dynamical systems (1-NFE) | RMFlow · Lorenz w/o event | TV distance: 0.0469 -> 0.0332 | Improved |
| add RMFlow refinement (KL, Lorenz w/o event) | RMFlow · Lorenz w/o event | KL divergence: 0.0109 -> 0.0059 | Improved |
| add likelihood loss weight lambda_1=1e-1 (vs 0) (ablation) | RMFlow · 2D checkerboard | KL divergence: 0.311 -> 0.163 | Improved |
| raise lambda_1 to 1e2 (ablation) | RMFlow · 2D checkerboard | KL divergence: 0.163 -> 0.425 | Hurt |
| raise lambda_1 to 1e1 (ablation) | RMFlow · QM9 | Molecule stability (%): 93.2 -> 83.9 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- StyleGAN-T — StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis (ICML 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.89)
- STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation (NeurIPS 2026 · similarity 0.88)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.88)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.88)
- TFG-Flow: Training-free Guidance in Multimodal Generative Flow (ICLR 2025 · similarity 0.87)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.87)
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.
