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

一句话要点

Policy-Mask 研究怎样从冻结的扩散模型里挑中间特征做判别任务：有用信号分散在去噪时间步、网络深度和通道上，它训练一个按预算条件的策略，在给定通道预算下联合选择时间步、模块和通道子集。

Policy-Mask selects intermediate features from a frozen diffusion model for discriminative tasks with a budget-conditioned policy trained on the downstream objective, jointly choosing timestep, block and channel subset under a prescribed channel budget instead of hand-fixed rules.