Research line · dominant
Flow matching and rectified flow
Regress the velocity field of a straight path between noise and data, rather than reversing a stochastic noising process.
What defines membership
The training objective should be a simple regression along a fixed, near-straight probability path.
How the line developed
Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.
The idea
Flow Matching · 2022-10landmark
A simulation-free objective for continuous normalizing flows, with the optimal-transport straight path as a simple special case.
Rectified Flow · 2022-09landmark
Independently derives the same idea, plus reflow to straighten paths further toward one-step sampling.
Evidence
SiT · 2024-01core
Places diffusion and flow matching in one interpolant family and shows the flow side wins at matched size and FLOPs.
At scale
SD3 / MMDiT · 2024-03landmark
Takes rectified flow to production scale with logit-normal timestep sampling; effectively settles the objective question for open text-to-image.
What it gets right
- Simpler to state, implement, and reason about than diffusion's schedule machinery
- Straighter paths make few-step sampling and distillation tractable
- Now the consensus choice, so tooling and intuition are mature
Where it is weak
- Gave up the well-developed SNR-weighting and noise-schedule theory of the DDPM era
- Path straightness in theory does not guarantee straightness in a given latent space
Reported numbers
As published, with the guidance method, budget, and model size that produced them. Seethe comparison page for why these cannot be ranked naively.
| Paper | Benchmark | Value | Guidance | Budget | Params |
|---|---|---|---|---|---|
| SD3 / MMDiT | GenEval | 0.74 | — | — | 8B + 4.7B T5-XXL |
| SiT | ImageNet 256x256 gFID | 8.3 | none | 7M steps | 675M |
| SiT | ImageNet 256x256 gFID | 2.06 | cfg=1.5 | 7M steps | 675M |
Also in this line
PixelFlow