MagicMatch Preview (GPU Full LUT)

MagicMatch Preview (GPU Full LUT) applies a LUT on the GPU using the cached feed tensor that MagicMatch Build LUT (GPU Full) stored in it - the preview half of the GPU Full pair. Standard inputs (source, lut, strength, and the three enums), an image output, WebGL live preview included. Because GPU Full keeps the parity-oriented hot path, this is the GPU preview that most closely matches what the CPU nodes produce - the main reason to pick it over the GPU Fast preview when your workflow has already committed to parity.

So how do you choose between the three GPU previews? If your Build came from the GPU Full node and you care about matching the CPU look, take this one - it keeps the parity-oriented apply path. If you just need the fastest feedback loop and can tolerate drift, GPU Fast preview is the snappier pair. The CUDA preview is the middle option: GPU build, CPU-parity apply.

The live-preview workflow is the same as the CPU Preview: queue once so Build runs and the LUT is cached, open this node to see the image under the widgets, drag strength (WebGL, instant, no re-queue), then queue once more before export. Missing feed tensor? It uploads the source itself and applies anyway, no error. render_mode default probe_export runs the GPU apply; the other modes fall back to CPU. The live-preview caveat from the CPU Preview still holds: the WebGL mid-strength view and the queued output can differ slightly, so that final queue is what makes the saved file match your slider.

Install: onnxruntime-gpu in ComfyUI's venv (uninstall onnxruntime first if it's present), restart, and the node appears under MAGICMATCH/GPU Full (experimental). Hide the trio with MAGICMATCH_GPU_FULL_NODES=0. Batch size 1.