PyTorch's missing toolbox for broken images and sparse sensors

What it does

deepinv is a PyTorch library that standardizes the plumbing for imaging inverse problems—think deblurring, MRI reconstruction, tomography, or super-resolution. It wraps physical forward models, reconstruction networks, training losses, and sampling methods into a single modular framework so researchers aren’t rebuilding the same blurry math for every paper. The library also ships with pretrained models and denoisers that work out of the box.

The interesting bit

The library treats imaging physics as first-class citizens: you get a catalog of predefined operators (the “how the image got mangled” part) paired with modern reconstruction strategies, including plug-and-play priors, unrolled optimization networks, and diffusion-based samplers for uncertainty quantification. That combination of physical modeling and deep learning in one toolbox is rarer than it should be.

Key highlights

• Predefined imaging operators for MRI, microscopy, tomography, deblurring, and more
• Pretrained reconstruction models and denoisers ready for immediate use
• Plug-and-play, optimization-based, and unfolded reconstruction frameworks
• Training losses and diffusion samplers tailored for inverse problems
• Official PyTorch Ecosystem project with active community maintenance

Verdict

Worth a look if you’re doing computational imaging research and tired of rewriting forward-model boilerplate. Probably overkill if you just need a generic image-to-image CNN and don’t care about the physics that corrupted your data.