Flow matching paper: velocity scaling cuts ImageNet-256 FID from 28.0 to 12.2, no retraining
francoisfleuret · x · 2026-10-14
A new arXiv paper by Youssef Saied and François Fleuret shows that scaling a trained flow-matching velocity field by a gain γ(t) — with no retraining — dramatically improves generation quality: on unguided SiT-XL/2 at 25 NFE on ImageNet-256, FID drops from 28.0 to 12.2 (the accompanying thread cites 8.7).
Key points:
- Refutes the prior claim that MSE training systematically underestimates velocity magnitude; the authors show no such deficit exists.
- Instead, velocity scaling reduces 'population time lag': sampled states at model time t resemble training states from an earlier time.
- Scaling velocity and rolling back model time are two equivalent ways to address this lag; measuring the lag to pick the gain improves quality across architectures and model sizes.
41 pages including appendix.
More from Multimodal
- Gemini 4 Argon builds an Elden Ring-style isometric voxel demo in one prompt — kastnerkyle · 2026-10-14
- Hollywood sold the dream of anyone making movies—until AI made it real — VraserX · 2026-10-14
- Gemini 4 Argon edits raw assembly to build a new level in a 36-year-old SNES game — venturetwins · 2026-10-14
- VigenFlow runs FLUX and Z-Image locally on AMD Ryzen AI NPU, no GPU needed — Kelsey_test · 2026-10-14
- You can generate custom voices right inside Google AI Studio — kastnerkyle · 2026-10-14
- Gemini 4 Argon scores a perfect mAP@50 100% on lego object detection test — zacharynado · 2026-10-14