The problem it solves
A classic denoising diffusion model is trained by adding Gaussian noise to an image according to a variance schedule, then learning to undo one increment of that at a time. Sampling walks that chain backwards. The trajectory is stochastic and curved, so each step has to be small, and quality only arrives after dozens to hundreds of network evaluations. The design also drags along a pile of choices that all interact: the beta schedule, whether the network predicts noise or velocity or the clean image, and which sampler you pair with which schedule.
Flow matching reframes the same goal. Pick a path between the noise distribution and the data distribution, then train the network to output the velocity that moves a sample along it. At inference you are no longer reversing a noising process, you are solving an ordinary differential equation from noise to image.
The simplest possible path is a straight line: mix noise and the real sample linearly, and make the training target the difference between them. That instantiation is rectified flow, and its virtue is entirely geometric. A straight trajectory can be traversed in large steps without leaving the path, so a solver that would overshoot on a curved route stays accurate on this one.
What changed in the models you use
Every recent frontier image and video family sits on this objective: SD3, Flux, the Wan video line, Hunyuan Video, Qwen-Image, LTX, and most of what shipped from 2024 onward. The user-visible consequences are specific.
Step counts collapsed. Twenty to thirty steps instead of fifty to a hundred. On video, where a single generation may involve tens of thousands of tokens per frame group, that difference is the gap between a viable product and an unaffordable one.
Distillation got easier. Straight paths are much simpler to compress into a handful of steps, which is why nearly every model family now ships a four-step or single-step sibling. The fast tier on a hosted catalog is usually a distilled version of the same flow matching base.
High resolution got more stable. These formulations shift the schedule based on how many tokens are being generated, because a large canvas needs more of its trajectory spent at high noise. That shift is why the same model can hold composition at 2K where the previous generation fell apart.
Guidance semantics changed. This is the one that trips people up. Several of these models expose a number labelled guidance that is not classifier-free guidance at all; it was baked in during distillation and is now a conditioning value. Habits from older models transfer badly, and the symptom is crushed contrast and banded colour.
Working with a flow matching model
A short practical checklist, from most to least useful:
- Start from the model's own recommended step count and guidance value. These are tuned per schedule and are not portable between families.
- If output looks washed out or flat at high resolution, that points at the schedule, not the prompt. Try the model's native resolution before rewriting anything.
- Treat the fast variant as a draft device. Explore composition on the distilled sibling, then run the final take on the full model with the same seed and prompt.
- Do not raise steps to fix detail. The detail ceiling is set by the latent encoder, and extra steps cannot add information the compression already discarded.
- When comparing two models, hold steps and guidance at each model's own defaults rather than matching numbers across them. Matched numbers on different schedules is not a fair test, and it is the single most common way model comparisons end up wrong.