MagicMatch one-shot (GPU Fast)

MagicMatch one-shot (GPU Fast) fuses the fast GPU pipeline into a single node: approximate scene extract, the ONNX color match, and the develop/LUT apply all run in one fused path on the device, with source, reference, and strength going in and a graded IMAGE coming out. It's the throughput pick of the one-shot family - the thing to reach for when you're processing many source images against a fixed reference and CPU build time is the thing standing between you and done.

The same shortcuts as the GPU Fast build apply - no JPEG normalize, bilinear reference resize, gray-world white balance - so parity with the CPU nodes is approximate by design. If that bothers you, the GPU Full one-shot keeps the parity-oriented pipeline at the cost of a heavier, more experimental code path; this node is the one that optimizes for getting frames through the queue. When render_mode is probe_export (the default) it runs the fused GPU apply; switch to polarr_probe or numpy_legacy and it falls back to the fast build plus a CPU apply instead. The other enums - lut_encoding (srgb_srgb) and profile_stage - behave exactly as they do across the pack.

Install: onnxruntime-gpu in ComfyUI's venv, after uninstalling the CPU onnxruntime wheel (they conflict on many builds):

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

Restart, and the node appears under MAGICMATCH/GPU Fast (experimental). Hide it with MAGICMATCH_GPU_FAST_NODES=0. Batch size 1. If the output looks slightly different from the CPU one-shot, that's not a bug - it's the speed-over-parity contract.