StyleGAN for Autolume

This branch (autolume) adapts NVlabs StyleGAN3 for use in Autolume, which builds on it for both model training and real-time inference. It stays close to upstream: changes are limited to packaging, Apple Silicon (MPS) support, the custom-ops build system and training quality of life. The main branch tracks NVlabs upstream, and the original StyleGAN3 documentation applies to everything not listed below.

Added features

Packaging

• uv-managed project: Python 3.12, torch 2.8.0 (cu128 index on Windows/Linux, PyPI on macOS), numpy 2.

Custom-ops build system

• Shared build cache keyed by source digest, torch version, CUDA version and compute capability: GPUs of the same capability share builds, and a warmed cache loads without a compiler, CUDA toolkit or ninja. Failed builds are detected and cleaned automatically.
• scripts/precompile_ops.py warms the same cache ahead of time, for the current device by default or an explicit --arch list.
• Cache entries record build provenance (torch, CUDA runtime, CUDA toolkit); builds warn on a toolkit/runtime major-version mismatch and precompilation refuses it unless --bypass-matching-toolkit is passed.
• CUDA 13 compatibility fix in filtered_lrelu.

Apple Silicon (MPS)

• --device option on gen_images.py, gen_video.py and train.py, defaulting to auto (cuda > mps > cpu). Ops without MPS kernels fall back to CPU.
• Training runs on MPS and CPU (single process, --gpus=1).
• The visualizer works on macOS: fixed-function imgui backend for the GL 2.1 context, 1:1 framebuffer, monitor work-area crash guard, and a working async renderer.
• Mouse presses are latched so short clicks register at low frame rates (all platforms).
• Texture mipmaps disabled on macOS (GL 3.0+ requirement).

Training

• resume_kimg is derived from the snapshot filename, so resumed trainings continue their kimg count and snapshot numbering.
• --augpipe presets from StyleGAN2-ADA (blit, geom, color, …, bgc default).
• Opt-in energy and emissions tracking via codecarbon (--track-emissions, off by default).

Dataset tool

• Palette (P-mode) images are converted to RGB instead of crashing the pipeline.

Quick start

git clone -b autolume git@github.com:Metacreation-Lab/stylegan3.git
cd stylegan3
uv sync
uv run python gen_images.py --network=<model.pkl> --seeds=0-3 --outdir=out

Optional, to ship or share prebuilt custom ops:

uv run python scripts/precompile_ops.py            # current GPU
uv run python scripts/precompile_ops.py --arch=7.5,8.6,8.9,12.0

License

NVIDIA Source Code License, unchanged from upstream — see LICENSE.txt.