Getting an LLM to Make Me a Tool for Enriching My Icons' Color Metadata

On my icon gallery sites, I have metadata I’ve manually added over the years to tag certain icons as being predominantly “blue” or “orange” or some other color.

Then I use this metadata to present icons of (roughly) the same color. It’s kinda neat to be able to browse a wall of icons that are all the same color.

The thing is: I know there are a lot of icons I’ve missed tagging over the years. But I have no idea how many, and figuring that out seems like a really arduous task. How do I go through 2,000+ icons and find all the ones that look predominantly “orange” but haven’t been tagged as such yet?

Seems like a good task to throw at an LLM. But I don’t want to just say, “Go tag everything that’s missing” and blindly trust the output. I need to be able to make a decisions as to whether I think a particular color is “orange” or not.

What I need is a tool for the job. I’m a very visual thinker, so to continue curating these color categorizations, I need some way for the computer to do its thing really, really fast, and then pull me into the loop to visually make decisions.

Here’s how I am thinking about about accomplishing this task:

That seems like it would be a good tool to put me in control of visual decision making around color categorization. So I tell the LLM to run with it.

We chat back and forth. I think, “You probably need to run all the icons through some model to make the correlation?” But it’s like, “Nah bro, just make a ‘hue histogram’.” It tells me how. For a color like “orange”, I can:

[process] each PNG, skip transparent pixels, skip low-saturation gray, convert the rest to HSV, and score how much of the remaining mass sits in the orange hue band (roughly 15–45°). Rank icons that don’t already have colorId: orange.

[This] scores each icon PNG by share of opaque pixels per color bucket, then writes a standalone HTML page: tagged vs maybe-missing, per color.

Let’s just make it, and then I’ll decide whether it’s good enough.

After a few iterations, the computer going “brr…”, and me saying “explain that like I’m dumb”, I have a really effective little tool!

The little threshold slider is a nice touch. It lets me fiddle around with the fidelity of the matches. In some cases, sliding it down reveals more icons I would’ve otherwise missed. In other cases, I’m like “What are you thinking? I don’t see that as ‘yellow’ at all!”

Supper effective little tool. I go through each color, select the ones I think are missing, paste the IDs back into the LLM, and then have it update each icon's metadata.

Boom, done! That all would’ve taken so long before. I would’ve never done it.

Reply via:

Email · Mastodon ·

Bluesky

Related posts linking here: (2026) VLM Enhanced Metadata For My Icon Galleries