Reproduce “Bayesian Tensor Decomposition with Diffusion Model Prior” 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
Zerui Tao, Qibin Zhao
Researchers from RIKEN AIP developed DiffBCP, a Bayesian tensor decomposition framework that integrates a low-rank CP prior with a learned prior from a diffusion model for robust signal recovery from incomplete and noisy observations. DiffBCP achieved up to 2.33 dB PSNR improvement over state-of-the-art baselines in image inpainting and denoising, while demonstrating strong generalization to out-of-distribution and high-resolution images.
Key insights
- Combining a structural low-rank prior with a learned diffusion model prior significantly enhances robustness in signal recovery from highly corrupted data.
- A Split Gibbs Sampler effectively decouples the complex posterior, enabling tractable inference by reformulating the diffusion prior step as a denoising problem.
- The CUSP prior and noise-adaptive coupling schedule allow for automatic rank determination and robust performance with reduced hyperparameter sensitivity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.