Color Correct GPU

If you've ever loaded a workflow off Civitai that finishes with Art Venture's Color Correct node, you know the controls: temperature, hue, brightness, contrast, saturation, gamma. They're handy, they look good, and for a single image they're fast enough that you never think about it. Then you batch that workflow across a hundred frames - or point it at an animation - and suddenly the whole thing crawls. That's the itch this node scratches.

Color Correct GPU is a from-scratch reimplementation of Art Venture's Color Correct with the same seven inputs, the same dial ranges, and the same processing order - but it does all the math with vectorized PyTorch ops on your GPU instead of looping frames through NumPy, PIL, and OpenCV on the CPU. The whole pitch is in one sentence of the README: same controls, same look, no per-frame Python round-trips.

Why you'd reach for it: whenever you're color-correcting a sequence rather than one image. Video pass-through, animation frames, or a big batch where you want the exact same grade applied to every output. On a single still it buys you almost nothing - this is a "when it matters, it matters a lot" node, not a daily driver. And there's a nice compatibility detail: its node ID is ColorCorrectGPU, so you can install it right next to comfyui-art-venture and both Color Correct and Color Correct GPU live in your menu without colliding.

How it works

The author deliberately kept the upstream's quirks so results stay close to the original. It preserves the 8-bit quantization steps and the PIL/OpenCV-style color math, just expressed as tensor ops. The order matches Art Venture: brightness, then contrast, then temperature, then gamma, then saturation (via HSL), then hue (via HSV). Neutral saturation and hue skip their color-space conversions entirely, so a mostly-default node is a cheap passthrough rather than a full round-trip.

Two implementation details worth knowing. First, if your IMAGE tensor is sitting on CPU, the node moves it to ComfyUI's torch device, does the work, then returns it to the intermediate device - so it plays nice with whatever pipeline it's dropped into. Second, big batches get chunked (~16M pixels at a time) so you don't blow up VRAM on an 8K batch pass. Don't expect byte-identical output to the original, though; the README is upfront that PIL/OpenCV and torch can differ on floating-point edge cases. Visually it matches, pixel-for-pixel it might not.

The inputs that matter

• image - your IMAGE tensor. That's the only input that isn't a slider.
• temperature - -100 to 100. Positive warms (pushes red), negative cools (pushes blue). Default 0.
• hue - -90 to 90. A true hue rotation in HSV space, not a tint. Default 0.
• brightness / contrast / saturation - -100 to 100, default 0. Same feel as the Art Venture versions: brightness scales, contrast pivots around mean luma, saturation scales in HSL.
• gamma - 0.2 to 2.2, default 1.0. Below 1 lifts shadows, above 1 crushes them.

Everything defaults to neutral, so the node is a no-op until you actually turn a dial - it slots into a workflow without changing anything. The single output is image, an IMAGE tensor, which wires straight into whatever comes next.

Installing it

ComfyUI Manager: search for ComfyUI-ColorCorrectGPU and hit install. Or, if you're doing it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/boobkake22/ComfyUI-ColorCorrectGPU

Then restart ComfyUI. The good news: there are no model files and no heavy dependencies. There's no requirements.txt at all - the only real requirement is torch, which ComfyUI already ships, plus Python 3.10+. It's a single 330-line file, which is exactly why it loads clean.

Gotchas

Honestly, there aren't many. The main trap is menu confusion: since the category is ArtVenture/Post Processing, if you also have Art Venture installed you'll see both Color Correct and Color Correct GPU in the same dropdown. Pick the one with the "GPU" - they're visually near-identical. Beyond that, if the node fails to load, it's a ComfyUI/torch mismatch rather than a missing package, and a manual clone fixes anything Manager chokes on. It's a small, focused, well-scoped tool. For batch color grading that matches the look you already know, it's the one you'd reach for.