The page presents Flow Map Trajectory Tilting (FMTT), a way to steer diffusion models toward samples that score well on a user-specified reward.
Approach: Instead of using an approximate denoiser to estimate the final sample, it works directly with a flow map, using its relationship to the velocity field.
Claims: The page says FMTT provably ascends the reward better than standard test-time methods that use the reward's gradient.
Uses: It supports exact sampling via importance weighting or search for local maximizers, and the page shows it enabling new image edits by interfacing with vision language models.
Sign in to comment.
TL;DR of the article, generated automatically.
The page presents Flow Map Trajectory Tilting (FMTT), a way to steer diffusion models toward samples that score well on a user-specified reward.
The paper is listed for NeurIPS 2026.