Reproduce “DPBridge: Latent Diffusion Bridge for Dense Prediction” 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

Haorui Ji, Taojun Lin, Hongdong Li

DPBridge introduces a latent diffusion bridge framework that adapts pre-trained diffusion models for efficient and high-quality dense prediction by operating from input images to target maps. This approach yields competitive performance on tasks like depth estimation, achieving an AbsRel of 4.4% on NYUv2, which represents a substantial reduction in error compared to previous diffusion bridge methods.

Key insights

  • Integrating pre-trained large vision models into diffusion bridge frameworks is essential to achieve robust performance in dense prediction, by leveraging extensive visual priors.
  • A "data-to-data" generative process, starting from the input image, is more efficient and conceptually suitable for dense prediction compared to traditional "noise-to-data" diffusion.
  • A maximum likelihood training objective, coupled with Distribution Alignment Normalization, enables effective fine-tuning of pre-trained diffusion backbones for bridge construction.

Summary by alphaXiv.

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