Rectified flow-matching inference with a FlowMatchEuler discrete scheduler
Use a scheduler that implements rectified-flow-matching sampling for latent diffusion inference. The approach is described as using a rectified flow objective to connect data and noise distributions along a straight-line path, paired with an Euler-style discrete step formulation. The scheduler also supports a resolution-dependent timestep shift parameter to better handle noise scaling at higher resolutions.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
The text describes that rectified flow-matching sampling can be simpler and can work well when reducing the number of sampling steps. It also provides practical controls (a scheduler choice plus a shift parameter) aimed at improving timestep/noise behavior across resolutions.
What changed
A new scheduler—FlowMatchEulerDiscreteScheduler—is introduced for SD3 inference. It adds (1) a rectified-flow-matching formulation with Euler method steps and (2) resolution-dependent shifting of the timestep schedule via a shift parameter.
Where it helps
- Text-to-image generation with SD3 using a Diffusers pipeline
- Controlling inference timestep behavior across different output resolutions
- Experimenting with fewer diffusion steps via rectified flow-matching sampling
- Swapping SD3 schedulers to use an Euler-based rectified-flow formulation
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The provided guidance only explicitly recommends a shift value for the 2B model (shift=3.0); the text does not give a general rule for other variants or how to select shift beyond that recommendation.
Sources
- Diffusers welcomes Stable Diffusion 3 ↗
Hugging Face · primary · Published 2024-06-12 · Checked 2026-11-10
