CAT-Flow: Training-Free Adaptive Steps Cut Flow Matching Generation by Up to 40%

chaumian · x · 2026-09-04

The arXiv paper "CAT-Flow" tackles the 20-30 step sampling bottleneck of Flow Matching (powering FLUX and Stable Diffusion 3.5) with two lightweight, training-free algorithms that adapt step sizes at inference time. Exploiting a connection between sampling ODEs and gradient flow, CAT-OT estimates curvature via finite differences of the vector field's time derivative while CAT-OV uses gradients over state space. Across four text-to-image Flow Matching models, they cut steps to comparable quality by up to 40%, outperforming existing step-size heuristics with constant-order truncation error bounds.

Original post →

More from Multimodal

Multimodal channel →