3 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Jul 7, 2026

Rectified flow is the quiet idea behind Stable Diffusion 3 and FLUX - and it is why modern image models are getting so fast.

Classic diffusion models are beautiful but slow. They reverse a noising process along a highly curved path, and following a curve accurately means many tiny steps - often 25 to 1000+ passes through a big network for one image.

Rectified flow asks a simpler question: what if the path from noise to data were a straight line?

Here is the whole idea:

1. Draw a straight line between a noise point and a real data point.
2. Interpolate a point on it: x_t = (1 - t) * noise + t * data.
3. Train a network to predict the velocity along it. The target is just data minus noise, learned with a plain mean-squared-error loss.
4. To sample, integrate that velocity from noise with a simple Euler solver - and because the path is nearly straight, a few steps is enough.

Then comes reflow: retrain the flow on pairs it generated itself to straighten the trajectories even more, and a distillation step can collapse it to one step. That is how InstaFlow turned Stable Diffusion into a one-step model at FID 23.3, generating an image in about 0.09 seconds - roughly 90% faster.

Straight path in, fast samples out - the recipe now powering SD3 and FLUX.

Did this make rectified flow click? Follow for plain-English GenAI breakdowns.

#GenerativeAI #DiffusionModels #RectifiedFlow #MachineLearning #AI