Research map / Image generation / Fast diffusion sampling
Mean Flow Distillation: Robust and Stable Distillation for Flow Matching Models
An Zhao, Shengyuan Zhang, Zhongjian Sun, Yixiang Zhou, Zejian Li, Ling Yang, Tianrun Chen, Lingyun Sun
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 |
|---|---|---|---|
| match time-integrated mean flow instead of instantaneous score (VSD / Diff-Instruct) | single-step student of OccFM · nuScenes 4D occupancy forecasting | mIoU (%): 26.42 -> 27.25 | Improved |
| distill OccFM into one step with MFD versus teacher at 1 NFE | OccFM · nuScenes 4D occupancy forecasting | mIoU (%): 8.15 -> 27.25 | Improved |
| use MFD instead of AYF consistency distillation | single-step student of OccFM · nuScenes 4D occupancy forecasting | mIoU (%): 21.69 -> 27.25 | Improved |
| use MFD instead of DMD2/SenseFlow (adversarial) distillation | 4-step SANA 1.6B · text-to-image, FID (DINOv2) | FID: 36.13 -> 33.28 (vs DMD2) | Improved |
| use MFD instead of AYF | 4-step SANA 1.6B · text-to-image | FID (DINOv2): 43.68 -> 33.28 | Improved |
| reduce ODE solver step size to 0.001 (ablation) | MFD · nuScenes 4D occupancy forecasting | mIoU (%): 27.25 -> 25.91 | Hurt |
| increase ODE solver step size to 0.1 or 0.5 (ablation) | MFD · nuScenes 4D occupancy forecasting | mIoU (%): 27.25 -> 26.47 (0.1), 26.54 (0.5) | Hurt |
| use moderate ODE step size 0.005 to 0.02 (ablation) | MFD · nuScenes 4D occupancy forecasting | mIoU (%): 27.14 (0.005), 27.25 (0.02) | Improved |
| temporal integration in MFD instead of VSD single-timestep gradient | MFD vs Diff-Instruct · text-to-image, SANA 1.6B | loss variance | Improved |
| reduce integration to single step (step size equals t-s) (ablation) | MFD · variance/stability analysis | training stability | Hurt |
| MFD training cost versus DMD2 | SANA 1.6B distillation · training cost | peak memory (GB): ~50 -> ~20 | Improved |
| MFD without adversarial or perceptual losses | MFD · text-to-image, SANA 1.6B | PickScore / HPSv2: 0.2002 vs 0.2004 (DMD2); 0.2442 vs 0.2445 (SenseFlow) | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Diff-Instruct — Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models (NeurIPS 2023) · better on 1 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
- Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.90)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.89)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.89)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.89)
- Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
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.
