MagicMatch Build LUT (GPU Full)

If GPU Fast is the node that trades parity for speed, MagicMatch Build LUT (GPU Full) is the one that wants both. It keeps the complete hot path on the GPU - detection buffers, luminance stats, the face ONNX preprocess, the develop@1600 render, the full-res LUT apply, even the net reference resize - with only minimal CPU sync left: the JPEG q98 normalize, detection-sized downloads, the 256×256 reference round-trips, and a single .cpu().numpy() feed into ONNX. Same source/reference inputs, same lut output, and it caches a GPU feed tensor for its matching Preview, just like GPU Fast.

In practice that makes it the one I'd reach for when I want the GPU's speed but also want to trust the result matches what the CPU nodes produce - the parity stack this pack keeps as its golden path. To be blunt about the lineup: GPU Fast build is for when you're iterating fast and accept drift; this node is for when you want the same parity commitment as CPU but faster. It's marked experimental, and of the three GPU flavors it's the heaviest, so it's the most code to have gone wrong. But the design goal is explicit: hot path on device, output close to CPU.

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

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

Restart, and the trio lands under MAGICMATCH/GPU Full (experimental). Set MAGICMATCH_GPU_FULL_NODES=0 before starting ComfyUI to hide them if you're not using them. Batch size 1, as everywhere in this pack.