Differentiable Photon Transport is a series about building a differentiable X-ray renderer in CUDA and Warp. We start with a ray passing through matter, work out what a detector measures, and learn how to differentiate the calculation. Those derivatives let us work backwards from an image to the geometry and material that produced it.

Three volumes, one renderer

Volume I: Foundations develops the physics, geometry and numerical methods, from primary transmission to scattering and differentiable transport. It closes with applications to registration, reconstruction and acquisition design.
Volume II: Neural and learned methods will cover neural representations and learned methods for photon transport and inverse imaging.
Volume III: Volume reconstruction will take the inverse problem further, using the forward model and its derivatives to recover three-dimensional material distributions from X-ray measurements.

By the end of the series, you will have built what amounts to Mitsuba for X-ray photons: a differentiable renderer that can simulate an acquisition and work backwards from its images.

A book with executable source

The code listings come directly from the executable implementation. You can follow a derivation into the kernel that implements it, run the accompanying example and inspect the result. Interactive figures let you explore the geometry and recorded calculations as you read.

About the author

I’m Chris von Csefalvay, a computational epidemiologist and AI researcher. As a Principal in HCLTech’s AI Practice, I lead research on post-training and clinical intelligence. I work on reinforcement learning from verifiable rewards, preference optimisation and agentic systems, with a particular interest in getting models to work reliably beyond the demo.

My work also extends to physical AI and robotics simulation. With Tamás Földi, I’m a principal investigator of the Asset Factory Blueprint, which turns photographs, scans and CAD into reproducible OpenUSD assets and simulation environments. I’m also one of the creators of CERES, which uses Meta Quest headsets to capture egocentric video and hand motion for research.

You can find my projects on GitHub and my research and writing at chrisvoncsefalvay.com. My other books are The Craft of Post-Training (No Starch Press, 2026) and Computational Modeling of Infectious Disease (Elsevier, 2023).

Corrections and questions

If you find an error, a broken figure or an explanation that does not make sense, please open an issue on GitHub. A link to the page and a short description are enough. For a code bug, include a small example if you can.