PSDR-CUDA is a physics-based differenetiable renderer written in C++17/CUDA and based on the theoretical framework introduced in the paper titled “Path-Space Differentiable Rendering”.
PSDR-CUDA is light-weight (compared to other modern renderers like Mitsuba 2) yet powerful, offering the following features:
1. Fast GPU-based ray tracing: PSDR-CUDA uses OptiX 7.1 that leverages RTX ray tracing on modern Nvidia graphics hardware.
2. Vectorized computation: PSDR-CUDA uses Enoki, the numerical foundation of Mitsuba 2, to perform vectorized computations (including reverse-mode automatic differentiation) on the GPU using CUDA 11.
3. Fast and unbiased gradients: PSDR-CUDA implements state-of-the-art algorithms that produce unbiased gradient estimates with respect to arbitrary scene parameters (e.g., spatially varying reflectance and object geometry).
4. Python bindings: PSDR-CUDA provides fine-grained Python bindings using pybind11 and can be easily integrated into Python-based interference and learning pipelines.
This project was created by Shuang Zhao. Significant features and/or improvements to the code were contributed by Kai Yan.