Reproduce “Generative Translation Priors: Bayesian Imaging with Cross-Modality Image Translation” 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

Evan Bell, Jiaming Liu, Yifan Chen, Yu Sun

Generative Translation Priors (GTP) turns cross-modality diffusion translation models into Bayesian priors for ill-posed image reconstruction, combining learned source-to-target relationships with explicit measurement likelihoods. The resulting Prox-GTP method improves severely undersampled CT reconstruction by up to about 7.6 dB over a diffusion baseline and yields higher PET reconstruction quality at moderate and high count levels.

Key insights

  • Cross-modality translation is used as a probabilistic prior rather than only as an image-generation or preprocessing module, allowing measured target-domain data to constrain the reconstruction.
  • The accumulated intrinsic bias tracks posterior-quality metrics in a Gaussian test problem, identifying an appropriate guidance regime without access to ground truth.
  • Side-modality information is most useful when the target measurements are highly incomplete: MRI guidance helps sparse-view and limited-angle CT substantially, while its relative benefit decreases as CT coverage improves.

Summary by alphaXiv.

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