comfyui-palette-toolkit

ComfyUI Palette Toolkit

Model-independent color-palette utilities for ComfyUI.

Text encoders do not reliably understand values such as #66788C. This project turns hex palettes into actual pixels and perceptual image operations instead of merely translating the codes into color names.

Initial nodes

Hex Palette Builder

Accepts a list of RGB hex codes and returns:

• a true RGB palette swatch as a ComfyUI IMAGE;
• normalized #RRGGBB text;
• JSON metadata for API workflows.

Palette Harmonizer

Pulls an input image toward the supplied palette using OKLab perceptual color distance. It supports soft attraction, chroma-only grading, and nearest-color quantization while optionally preserving source luminance.

Because the harmonizer consumes and emits ordinary images, it can be used with any model family:

Load Image → Palette Harmonizer → VAE Encode → I2I sampling

or:

VAE Decode → Palette Harmonizer → Save Image

The first arrangement lets the palette influence generation through the model's image pathway. The second provides deterministic finishing. A model-specific IPAdapter or style-reference path can also consume the palette image, but that integration is deliberately outside the core model-independent nodes.

Installation

Clone or copy this repository into ComfyUI/custom_nodes/ComfyUI-Palette-Toolkit, then restart ComfyUI. No additional package installation is required; the node uses the PyTorch already supplied by ComfyUI.

First experiment

Use a six-color palette, soft mode, 0.35 strength, 0.08 softness, and preserve luminance. Grade the I2I source before VAE Encode, then compare an unchanged seed at several denoise strengths.

Development

Dependency-free parser tests can be run from the repository root:

python -m unittest discover -s tests -v

Runtime image tests require the Python environment used by ComfyUI because the color processing operates on its PyTorch tensors.

Roadmap

• Generate palette maps with configurable spatial color proportions.
• Extract dominant palettes from reference images.
• Add masks and per-region palette assignments.
• Produce and apply reusable 3D LUTs.
• Add optional bridges for model-specific visual-conditioning systems.
• Explore differentiable palette guidance during sampling.

Design boundary

This project provides genuine pixel-level palette control. It does not claim that one universal conditioning object can plug directly into every diffusion architecture; those adapters remain model-specific.

License

MIT