Perceptual Flow Matching for Accelerated Few-Step Generation
July 2, 2026
Perceptual Flow Matching (PFM) supervises flow matching in a pretrained perceptual feature space rather than standard VAE latent space. This reduces required sampling steps from 35–50 down to 4–8 without requiring teacher models or auxiliary score networks.
HOW THIS AFFECTS YOU
●
builderYou can drastically accelerate generative model inference with minimal changes to the training pipeline.
●
researcherThis offers a way to achieve few-step generation without the complexity of distillation or teacher-student setups.