MagicMatch Preview (CUDA LUT)

MagicMatch Preview (CUDA LUT) is the Preview node paired with MagicMatch Build LUT (CUDA). The idea is clean: build the LUT on the GPU where it's fast, but apply it through the full-res CPU parity path so the final image matches what the CPU nodes would have produced. Inputs are the standard Preview set - source, lut, strength, plus lut_encoding / render_mode / profile_stage - with an image output, and you get the same in-node WebGL live preview as the plain Preview: run once, open the node, drag the slider, queue again before export.

The one behavioral switch is inherited from the CUDA one-shot. Set MAGICMATCH_CUDA_GPU_APPLY=1 and the probe_export apply routes through the legacy GPU path, which may come out brighter. Default 0 keeps CPU parity. Wire it after Build CUDA for the intended pairing - though any LUT flows into it, since they're all the same custom type, the caching and parity story is only coherent when the LUT came from the CUDA build. Note the scope, too: only the ONNX color-match build gets the GPU here. Detection, develop, and apply stay CPU, which is exactly why the output tracks the CPU nodes so closely.

The live-preview workflow is identical to the CPU Preview: run once, open the node, drag strength on the WebGL preview, then queue once more so the queued image matches your slider before you export.

Install is the CUDA flavor of the pack - in ComfyUI's venv, uninstall onnxruntime, then:

pip install -r MAGICMATCH/requirements-cuda.txt

Restart, and the node shows up under MAGICMATCH/CUDA (experimental). No CUDA on the box means it quietly falls back to CPU execution. Batch size 1.