MagicMatch Preview (GPU Fast LUT)

MagicMatch Preview (GPU Fast LUT) is the apply-side partner to MagicMatch Build LUT (GPU Fast), and the two are designed to be used together: the build caches a GPU feed tensor inside the LUT, and this Preview reuses it for a fast GPU develop + LUT apply instead of re-uploading the source. Inputs are the standard set - source, lut, strength, lut_encoding, render_mode, profile_stage - and you get the WebGL live preview along with the image output.

If you wire in a LUT that has no cached feed tensor (say, from a CPU Build), it just uploads the source and applies - no error, you only lose the caching speedup. And consistent with GPU Fast's design, the apply is the fast GPU path, so parity with the CPU Preview is approximate by design. render_mode is probe_export by default; the other modes fall back to a CPU apply.

The workflow is the same as the CPU Preview, just faster: queue once so the Build runs and caches, open this node and drag strength on the live WebGL preview, then queue once more to make the image output match before you save. Choose this preview over the GPU Full one when the loop matters more than bit-identical parity; choose the GPU Full preview when you want the GPU speed without accepting drift.

Install: onnxruntime-gpu in ComfyUI's venv (uninstall the CPU onnxruntime wheel first - they conflict on many builds), restart, and the node appears under MAGICMATCH/GPU Fast (experimental). Hide with MAGICMATCH_GPU_FAST_NODES=0. Batch size 1. And the same live-preview note as the CPU Preview: mid-strength WebGL preview and the queued output can differ slightly, so queue once more before you export.