nanodrr

A performance-oriented reimplementation of DiffDRR with the following improvements:

• Optimized, pure PyTorch implementation (~5× faster than DiffDRR at baseline)
• Modular design (freely swap subjects, extrinsics, and intrinsics during rendering)
• Compatibility with torch.compile and mixed precision
• Extensive type hints with jaxtyping
• Standard Python package structure managed with uv

All projective geometry is implemented internally using the standard Hartley and Zisserman pinhole camera formulation.

Installation

Note

On pytorch<2.9, torch.compile with bfloat16 is slower than eager due to a CUDA graph capture issue (see Benchmarks). Use pytorch>=2.9 (Triton ≥3.5) for best results.

To strictly install the renderer:

pip install nanodrr

To install the optional 3D visualization module:

pip install "nanodrr[scene]"

Benchmarks

Important

• ~5× faster than DiffDRR out of the box, without compilation (946 FPS vs 213 FPS)
• ~8× faster with torch.compile and bfloat16 on pytorch>=2.9 (1,650 FPS vs 213 FPS)
• ~2.5× less memory than DiffDRR (516 MB vs 1,344 MB peak reserved with bfloat16 + compile)

Mean ± std. dev. of 10 runs, 100 loops each. Benchmarked by rendering 200×200 DRRs on an NVIDIA RTX 6000 Ada (48 GB) with Python 3.12. Compile represents torch.compile(mode="reduce-overhead", fullgraph=True). Full experiment at tests/benchmark/.

Docs

To test the docs locally, run

uv run --group docs jupyter nbconvert --to markdown tutorials/*.ipynb --output-dir docs/tutorials/
uv run --group docs zensical serve

Roadmap

• Implement a fully optimized renderer
• Port strictly necessary modules from DiffDRR (e.g., SE(3) utilities, loss functions, and 2D plotting)
• Migrate 3D plotting functions to an optional module
• Integrate with xvr to speed up network training and registration
• Integrate with polypose to speed up registration
• Release as v1.0.0 of DiffDRR!