MagicMatch one-shot (CUDA)

MagicMatch one-shot (CUDA) is MagicMatch with the ONNX build moved onto the GPU and everything else left alone. Same shape: source + reference + strength in, graded IMAGE out, plus the lut_encoding / render_mode / profile_stage enums defaulting to srgb_srgb / probe_export / current_profile_stages. The apply side stays on the CPU parity path by default, which is why the output should match the plain one-shot - the GPU only speeds up the part that builds the LUT.

There's one switch worth knowing about. Set MAGICMATCH_CUDA_GPU_APPLY=1 and the apply also moves to the legacy GPU path, which the README warns "may look brighter." Unless you're chasing that specific behavior, leave it at 0. The env-var hiding works the same as the other CUDA nodes: MAGICMATCH_CUDA_NODES=0.

Because it's a one-shot, the CPU-node trade-off still applies: every strength change re-runs the build. On CUDA that's faster, but it's not free, and if you're going to be dragging the slider, the Build + Preview pairing is still the smoother loop. This node is the right pick when you want the all-in-one convenience of MagicMatch but your pairs are big enough that the ONNX build time actually hurts.

Install is the CUDA flavor of the pack - in ComfyUI's venv:

pip uninstall -y onnxruntime
pip install -r MAGICMATCH/requirements-cuda.txt

Restart, and it shows up under MAGICMATCH/CUDA (experimental). The README's focus is H100/Linux, but the fallback behavior is forgiving: on machines without CUDA, the nodes drop to the CPU execution provider automatically. Batch size 1.