25 upvotes

Policy-Mask trims diffusion features for classification

Paper · Policy-Mask: Budget-Aware Adaptation of Diffusion Features to Discriminative Tasks

What is Policy-Mask: Budget-Aware Adaptation of Diffusion Features to Discriminative Tasks?

Policy-Mask automatically selects timesteps, layers and channel groups from a frozen diffusion model to use for discriminative tasks. It learns a lightweight, budget‑aware policy so you don’t have to hand‑pick features across the denoising trajectory.

Who it is for

Anyone repurposing big diffusion models for vision tasks can get strong performance with less compute and no extra training.

The numbers

Open Policy-Mask: Budget-Aware Adaptation of Diffusion Features to Discriminative Tasks →