MagicMatch Build LUT (GPU Fast)

MagicMatch Build LUT (GPU Fast) is the pack's "speed over parity" LUT builder. Same inputs - source, reference - same lut output, different priorities: it gets the LUT built as fast as the GPU can manage, and it's happy to take shortcuts the CPU path won't.

Three shortcuts, specifically. It skips the JPEG q98 normalize and uploads the Comfy IMAGE tensor straight to the GPU. It resizes the reference with bilinear instead of the WebP/Lanczos round-trip. And it does white balance with a GPU gray-world estimate rather than the face-aware model pass. Each one is a deliberate trade: faster, and occasionally a slightly different look than the CPU build. If you're iterating over many sources against one fixed reference on a tight loop, this is the node that exists for that.

The other thing it does is cache a GPU feed tensor inside the LUT it emits - which is why you pair it with MagicMatch Preview GPU Fast LUT. That Preview reuses the cached tensor instead of re-uploading, so the whole build + preview loop stays on device. Wire its LUT into the plain CPU Preview and it still works (the Preview falls back to a fresh upload), you just lose the caching bonus.

Install as with the other GPU variants: onnxruntime-gpu in ComfyUI's venv, after uninstalling the CPU onnxruntime wheel (they don't coexist on many builds):

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

Restart, and the nodes appear under MAGICMATCH/GPU Fast (experimental). Set MAGICMATCH_GPU_FAST_NODES=0 to hide them. Batch size 1, and expect parity differences by design here - this is the node you pick when throughput beats bit-identical output.