The paper introduces MrFlow, a training-free acceleration strategy for pretrained flow-matching text-to-image models. It generates global structure at low resolution, upsamples in pixel space with a lightweight pretrained GAN, injects low-strength noise, and resamples details at high resolution. On FLUX.1-dev and Qwen-Image, the authors report up to 10x end-to-end acceleration with OneIG within a 1% gap of the unaccelerated baseline. MrFlow is also described as orthogonal to pretrained timestep distillation, reaching up to 25x acceleration when combined. The reported method requires no additional training or runtime dynamic identification.
No heat snapshots are available in the last 24 hours.