☁️BizyAir Color Pallete

The node calls itself "Color Palette" in the UI and "ColorPreprocessor" in the class name, and both names are honest about what it does: it takes a normal image and reduces it down to a coarse grid of color blocks - basically a heavily pixelated, low-detail version - that captures the rough palette and tone layout of the source while throwing away everything about shapes, edges, and fine detail. That output isn't meant to be looked at; it's meant to be fed into a T2I-Adapter color model as the conditioning image. The category name in ComfyUI gives it away - this sits under BizyAir's "T2IAdapter-only" preprocessor group, not the general ControlNet one.

T2I-Adapter is worth knowing as its own thing, not just "ControlNet's cousin." It's TencentARC's lighter-weight conditioning framework - smaller, cheaper adapters than a full ControlNet, and it's still the pick when VRAM is genuinely the constraint. The color mode specifically is a narrow tool: unlike canny or depth conditioning, which lock down structure and composition, color conditioning only nudges the overall palette. Feed it a sunset photo and generate "a cat," and you'd expect a cat rendered in warm oranges and purples, while the model stays completely free to invent the cat's pose, the background, everything else. It's the tool you reach for when you want a mood or color scheme carried over from a reference without dragging its composition along with it.

Inputs that matter: just image, required - the source you're extracting a palette from. The only other control is resolution (optional, default 512, range 64–1024 in steps of 64), which sets the working resolution before the palette grid gets built. Because the whole point of this preprocessor is to produce something coarse, cranking resolution up doesn't buy you more detail in the useful sense - it mostly changes how finely the color regions get sampled before being blocked out. If your source image has small but distinct color regions you want preserved (a striped shirt, say), a higher resolution keeps those blocks separate instead of averaging them together; for a simple scene, the default is fine.

Output: IMAGE - the color-block map itself, ready to hand to a T2I-Adapter color model and its apply node elsewhere in your graph. Nothing BizyAir-specific about the type, so it composes with any T2I-Adapter setup you've already got, cloud or local.

Installing it: search "BizyAir" in ComfyUI Manager and install, git clone https://github.com/siliconflow/BizyAir.git into custom_nodes and restart, or comfy node install bizyair on the Comfy CLI - one install covers this preprocessor along with the rest of the pack. This node's own compute (turning your image into a color grid) is a genuinely light operation, but it still routes through BizyAir's API like everything else here, so you need a registered account and your API key set before it'll run.

Where people get tripped up: this is a niche, single-purpose tool, and the most common mistake is reaching for it expecting ControlNet-grade structural control - it won't lock in composition or pose, only palette, so if your actual goal is "make the output look like this reference," pair it with something structural (canny, depth) rather than relying on color alone. Beyond that, treat any stall or failure the way you'd treat any other BizyAir node: check that your API key is actually set before assuming the graph itself is broken.