Reproduce “Flow Priors for Linear Inverse Problems via Iterative Corrupted Trajectory Matching” 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

Yasi Zhang, Peiyu Yu, Yaxuan Zhu, Yingshan Chang, Feng Gao, Ying Nian Wu, Oscar Leong

Iterative Corrupted Trajectory Matching (ICTM) provides an efficient iterative algorithm for approximating the Maximum-A-Posteriori (MAP) estimator, leveraging flow-based generative models as priors for linear inverse problems. This method consistently achieves superior image reconstruction quality (e.g., higher PSNR and SSIM) across tasks like super-resolution, inpainting, and compressed sensing on natural and medical datasets, while improving computational efficiency.

Key insights

  • The global Maximum-A-Posteriori (MAP) objective can be effectively approximated by a weighted sum of local MAP objectives focusing on small segments of the generative flow's trajectory.
  • Efficient computation of the score function (gradient of log-likelihood) for flow priors can be achieved using Tweedie's formula, circumventing the need for expensive backpropagation through the entire ODE solver.
  • Aligning a corrupted data trajectory with the generative flow's path allows for a tractable and efficient local data fidelity term in the optimization process.
  • The algorithm demonstrates robustness to hyperparameter choices for step size and guidance weight, with K=1 iteration per timestep generally sufficient for natural image tasks, and K=10 beneficial for more complex problems like compressed sensing.
  • NFEs (N) around 100 offer a good balance between reconstruction quality and computational time.

Summary by alphaXiv.

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