Reproduce “DP-IQA: Utilizing Diffusion Prior for Blind Image Quality Assessment in the Wild” 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

Honghao Fu, Yufei Wang, Wenhan Yang, Alex C. Kot, Bihan Wen

DP-IQA repurposes the denoising U-Net of Stable Diffusion as a visual prior for blind image quality assessment, combining multi-scale diffusion features with text and image adapters to capture both semantic content and low-level distortions. Its teacher model reaches PLCC/SRCC scores of 0.913/0.893 on CLIVE, 0.951/0.942 on KonIQ, and 0.683/0.579 on LIVEFB, while distillation produces an 81-million-parameter student that runs about three times faster with modest accuracy loss.

Key insights

  • Diffusion features provide a stronger frozen representation for BIQA than the evaluated CLIP and ResNet-50 features, supporting the idea that denoising priors retain quality-relevant low-level information.
  • Both textual conditioning and direct image features contribute: removing either adapter lowers performance, and combining multiple U-Net resolutions is better than relying on a single feature level.
  • Feature distillation is important for the compact model, allowing it to approximate the teacher’s intermediate quality representation rather than learning only the final score.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.