ComfyUI Color Bias

Color Bias is a custom ComfyUI node that does two things at once:

1. Adds color-focused text conditioning to an existing CONDITIONING input.
2. Builds a palette-based LATENT that can be sent to KSampler.

The node is designed to steer color without requiring a separate image reference node. For models such as Anima, SD15, or SDXL (and variants of it - illustrious, noobai) that have a strong connection between color conditioning and the latent space, this can be a powerful way to achieve more vibrant and accurate colors in your generations.

Node

Display name: Color Bias

Outputs:

• conditioning: the original conditioning plus additional color conditioning.
• latent: a generated palette latent adapted to the connected model's latent format.

Required Inputs

conditioning

Base positive or negative conditioning that you want to modify.

model

clip

vae

bypass

Boolean switch.

• False: normal behavior.
• True: skips all color processing, returns the original conditioning unchanged, and returns a neutral non-colored latent.

active colors

How many color slots are active.

Valid range: 1 to 5.

Only the first N color inputs are used.

global influence

Overall strength of the added color conditioning.

• Lower values: weaker color push.
• Higher values: stronger color push.
• 0.0: keeps the original conditioning unchanged, but the node still returns the generated latent.
• 0.5: the default value, provides a nice balance in most cases.

batch size

How many palette samples to generate in one batch.

This affects the size of the returned latent batch.

width

Width of the generated palette image before VAE encoding.

Higher values increase cost and memory use.

height

Height of the generated palette image before VAE encoding.

Higher values increase cost and memory use.

latent

Palette generation mode if you need extra color conditioning.

Available modes:

• noise: creates a noisy palette made from the selected colors.
• vertical: creates vertical color bands. (experimental, may be less efficient than noise)
• horizontal: creates horizontal color bands. (experimental, may be less efficient than noise)

color 1,2,3,4,5

Color in hex form from the ComfyUI color picker.

strength 1,2,3,4,5

Relative weight of each color.

This affects both:

• how strongly the color contributes to the generated palette latent,
• how strongly the color contributes to the added color conditioning.

Output Details

conditioning

Returns a CONDITIONING object built from:

• the original input conditioning,
• one additional color-conditioning entry per active color.

Each added entry is generated from a color description derived from the hex value. The node tries to use webcolors first for exact named colors and falls back to hue-based naming if no exact CSS-style name is available.

latent

Returns a LATENT dictionary with encoded samples.

The latent is generated from a synthetic palette image and then adapted to the connected model's latent format. This is useful for models with non-standard latent channel counts.

If bypass=True, the returned latent is neutral gray rather than colorized.

Typical Workflow

1. Connect your positive conditioning into conditioning.
2. Connect the same model, clip, and vae that you use in the rest of your workflow.
3. Pick 1 to 5 colors.
4. Adjust strength values to control the palette balance.
5. Send the node's conditioning output into the positive input of KSampler.
6. Send the node's latent output into the latent_image input of KSampler.

Performance Notes

• Large width and height values make the node slower.
• Larger batch size increases memory use and processing time.
• More active colors means more CLIP conditioning work.
• noise mode is heavier than simple banded layouts, even with optimization.

Installation

Git clone this repository to ComfyUI custom_nodes folder, then restart ComfyUI.

This node depends on:

• webcolors