Diffusion Models

The research presents Policy-Mask, a framework designed to efficiently adapt frozen diffusion features for downstream discriminative tasks under strict computational budgets. It achieves this by learning a policy to select optimal feature channels and layers through reinforcement learning.

The paper proposes DaTMe, a new method for accelerating diffusion model inference through intelligent token merging. Unlike previous methods that focus only on text-aligned regions, DaTMe preserves important structural and temporal details by analyzing the underlying denoising dynamics.

The researchers present a novel approach to reinforcement learning for image generation by repurposing pretrained MLLMs as zero-shot reward models. By measuring how well a prompt can be recovered from a generated image, the method avoids the need for costly human preference labeling.