devzhk/InverseBench

Benchmarks plug-and-play diffusion priors across five scientific inverse problems (optical tomography, medical imaging, black hole imaging, seismology, fluid dynamics) with domain-specific evaluation metrics and 14 algorithm implementations. Uses a modular, Hydra-config-driven architecture with composable forward operators, diffusion priors, algorithms, and evaluators that interact through a unified pipeline. Includes pre-trained diffusion models for each domain, comprehensive hyperparameter search infrastructure via Weights & Biases, and Docker/UV environments supporting specialized dependencies like EHTIM for black hole imaging and Devito for seismic simulation.

105 stars.