Badge Color ID Map V1

What this is for

Here's the problem this node solves. You want to select "the teal region" of a rendered badge. In the flat artwork it was one hex value. In the render, after the model has done its lighting, that teal is a hundred different RGB values across a gradient - and your colour picker only ever catches one of them. Eyeballing a tolerance gets you a mask with holes in it.

Badge Color ID Map V1 asks an image model to strip the appearance out: same canvas, same geometry, same every-line-and-glyph, but each closed region repainted in one flat palette colour from a fixed 16-colour set. Then the picking is trivial and exact, because "the teal region" is now literally #00B0FF everywhere. The pack's own framing is the right one - the map never replaces your edit image, it's a lookup table for where things are.

Cost warning up front: this is a paid generation. It goes through the official ComfyUI API node path and bills against your account credits. Which is why the node has a cache, and why you should think of this as a once-per-design expense.

How it works

When enabled is true it validates that the incoming image is exactly one 1024×1024 RGB image - the badge pipeline's canvas is fixed, so this is a hard requirement, not a suggestion. Then it builds a cache key from everything that affects the output: the master image digest, prompt version, the prompt text, model, size, background mode, quality, seed and map_revision.

Cache hit, and it loads the stored PNG and returns immediately with cache: hit. Cache miss, and it expands a subgraph containing a native OpenAIGPTImageNodeV2 - officially a paid API call - with your prompt, model, 1024×1024 opaque output and quality setting, and the result is what gets cached.

The map_prompt default is worth reading rather than replacing. It instructs the model to preserve canvas, composition, geometry, topology, outlines, separator lines, text glyphs and spacing, treats lighting/reflections/gradients as appearance rather than new regions, forbids merging or redrawing regions, restricts output to that 16-colour palette, reserves black for background and genuine holes, and demands no legend, no shading, no new outlines. It's a well-tuned prompt and the short version of the whole node.

enabled false is a no-op: you get a 1×1 placeholder image, an empty cache_key and state: not_generated. The lazy input means the master image isn't even evaluated when it's off, so a disabled branch costs nothing.

The inputs and outputs

• enabled - off by default. Turn it on when you actually need a map.
• master_image - one 1024×1024 render. Anything else raises an error about the required canvas.
• map_prompt - the instruction text. Keep the default unless you have a specific reason; if you do edit it, the cache key changes and you pay again.
• quality - low / medium / high, default high.
• seed, map_revision - seed for the generation, and a manual counter to force a fresh map when you changed something the key can't see.
• model (optional) - gpt-image-2 (default) or gpt-image-2.5-sunburst.

Outputs: color_id_map (the flat map), cache_key (feed it to the cache store or a downstream picker), status (JSON with state, cache hit/stored, and the key).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/TeutonicMepheles/ComfyUI-DAELab-Custom-Nodes-Library.git
# restart ComfyUI

This node is a V3 node, so it needs a current ComfyUI core with comfy_api.latest. If Badge Color ID Map V1 simply doesn't appear in the node list while the rest of the pack does, that's the reason - update ComfyUI rather than pip-installing anything called comfy_api.

Where people get burned

Palette collisions. Sixteen colours for every closed region means a very busy badge will run out. When the model reuses a colour on two adjacent-but-should-be-different regions, your later selection silently covers both. Split the design or accept a coarser region set.

Model creativity. It's a generative call, not a filter. Occasionally it smooths a hairline or thinks a stray highlight is a region. The diagnostics in the pack's own docs are honest that thin gold lines and weak boundaries can come back imperfect; a map_revision bump gives you a re-roll for one more paid call.

Non-square inputs. The map path wants 1024×1024, so a non-square render has to be contain-fitted upstream (the pack's stage does the padding and the nearest-neighbour un-padding for you) - feeding it a 1024×640 will just error, not crop.

Forgetting it's cached per-model. Switching between Sunburst and plain GPT-Image-2 changes the key, so old maps won't be reused. That's correct behaviour and it still surprises people mid-session.