Flow Matching revisited: the simulation-free CNF training method behind today's generative models
burny_tech · x · 2026-09-21
A thread revisits the 2023 Meta paper "Flow Matching for Generative Modeling" (Lipman, Chen et al., arXiv:2210.02747), which generalizes score-based generative models via continuous normalizing flows. Key points: Flow Matching trains CNFs simulation-free by regressing vector fields of fixed conditional probability paths; diffusion paths are just a special case of its Gaussian path family, and FM makes training more stable; optimal-transport interpolation paths yield faster training/sampling and better generalization; on ImageNet, FM-trained CNFs beat diffusion baselines on both likelihood and sample quality. It underpins today's rectified-flow generative models.
More from Multimodal
- Meme: GPT-Image killed the PowerPoint star — cocktailpeanut · 2026-09-21
- Runway CEO teases announcement: the question he's been asked for five years gets answered tomorrow — c_valenzuelab · 2026-09-21
- Redditor finds no SynthID watermark in dozens of Qwen 2.1 images — Bulky-Employer-1191 · 2026-09-21
- GLSL shader workaround removes grid noise artifacts in Qwen Image 2.1 — thevegit0 · 2026-09-21
- Qwen Image Edit quick samples show off editing results — malcolmrey · 2026-09-21
- Reddit thread: keeping facial identity and likeness consistent in Stable Diffusion — Opposite_Yam_4161 · 2026-09-21