Reproduce “Direct Diffusion Score Preference Optimization via Stepwise Contrastive Policy-Pair Supervision” 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
Dohyun Kim, Seungwoo Lyu, Seung Wook Kim, Paul Hongsuck Seo
DDSPO aligns diffusion image generators by comparing preferred and dispreferred denoising behavior at intermediate steps, using contrastive-policy score targets rather than deriving preference targets only from finished images. It can use labeled preferences or original and degraded prompts, though training still uses forward-noised inputs. On SD-1.4, GenEval rises from 0.4245 to 0.5045, but DDSPO does not lead every benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.