Reproduce “DRM: Diffusion-based Reward Model With Step-wise Guidance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaxon Zhang, Binxin Yang, Hubery Yin, Chen Li, Jing Lyu
DRM adapts a pretrained SD3.5-Medium diffusion transformer into a reward model that scores noisy intermediate image states. Step-GRPO uses these local scores during training, while Step-wise Sampling branches and selects during inference. On SD3.5-Medium, DRM with Step-GRPO scores 10.28 on HPSv3 versus 10.07 with standard GRPO; the comparison does not isolate feedback from changed group aggregation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.