AI
PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 16:00:41 UTC
Briefing created: 01/10/2026, 09:02:13 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 22:18:24 UTC.
Briefings
PreviewDiff is a training-free test-time search method for diffusion models that uses multimodal feedback to guide generation during the denoising process. By decoding partial previews at selected checkpoints, the system obtains natural-language critiques from judges, which are then used to branch over semantic prompt edits and locally re-noise latent continuations. This approach improves over budget-matched Best-of-N selection and strong scalar-search baselines across image and video generation benchmarks, demonstrating that multimodal feedback acts as an active controller rather than just a final verifier.
Key points
- PreviewDiff improves over Best-of-N selection and scalar-search baselines across image and video generation benchmarks.
- The method uses multimodal feedback to branch over semantic prompt edits and locally re-noise latent continuations during denoising.
Context and limitations
Coverage is limited to image and video generation benchmarks; results are from a single snapshot on 2026-09-28 without long-term stability data.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 09:02:13 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigjhi55akwnc7vqy4k624tx6n7ege5u3jcjm2tmtevsluleacc5em