Flower: A Flow Matching Solver for Inverse Problems

by Mehrsa Pourya (EPFL)

Thursday 20 Aug 2026, 11:30 → 13:00 Europe/Zurich

OHSA/B17

Abstract:
Inverse problems involve the reconstruction of a signal from incomplete or noisy measurements. Image restoration tasks, such as super-resolution, deblurring, and inpainting, are inverse problems. Their solution needs a prior on what natural images look like. Flow matching models are the current state of the art for image generation and provide exactly the kind of prior required. In this talk, I will present Flower, a method that transforms a pretrained flow matching model into a solver for inverse problems by modifying its inference path. I will outline how Flower works, show results for different tasks, and discuss future work.

TEAMS link here