PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents
PreviewDiff steers diffusion runs before they harden into bad final samples.
The paper describes a training-free test-time search method that checks intermediate denoising previews with a multimodal judge. Its feedback is used to branch through prompt edits and locally re-noised latent continuations, then keep rolling forward the better branches. The authors report gains over budget-matched Best-of-N and scalar-search baselines on image and video benchmarks. Their ablations say earlier interventions and wider search help most, with deeper search and extra semantic variants adding further gains. HF Daily Papers' note
The paper describes a training-free test-time search method that checks intermediate denoising previews with a multimodal judge. Its feedback is used to branch through prompt edits and locally re-noised latent continuations, then keep rolling forward the better branches. The authors report gains over budget-matched Best-of-N and scalar-search baselines on image and video benchmarks. Their ablations say earlier interventions and wider search help most, with deeper search and extra semantic variants adding further gains. HF Daily Papers' note
score 5