13,000+ icons, generated by AI: what I actually learned building Thridy
When I started Thridy, I thought the hard part would be the icons. Generate something that looks good in soft 3D, ship it, repeat a few thousand times. I was wrong about almost everything.
I should say this plainly up front, because it shapes everything below: the icons are generated with AI. I direct the style, curate hard, clean up the output, and do all the work around them, but the images themselves come out of a model, not out of me hand-modeling 2,600 objects in a render engine. Generating them turned out to be the easy part. The real work was in all the decisions around it: how to keep 2,600 AI-generated icons looking like they came from the same hand, how to name them so people can actually find them, and how to make a soft little render feel useful in a real product instead of decorative.
Here is what I learned, honestly, including the things I got wrong.
Consistency is a system, not a vibe
Early on I treated each icon as its own generation. I'd write a prompt, take whatever came back that looked nice that day, and move on. Then I put twenty of them next to each other and the whole thing fell apart. One was lit from the left, another from the top. One sat at a slightly steeper angle. Two greens that should have matched did not. Individually every icon was fine. Together they looked like a pile of stock assets from different sellers, which, in a sense, is exactly what unconstrained AI output is.
That is the thing about generating with a model: every image is an independent roll of the dice. Left alone, it will not give you a family. It will give you 2,600 cousins who have never met.
The fix was boring and it was the most important thing I did. I locked the style. One fixed reference for the camera angle, the light direction, the surface finish, and the framing, a scaffold I do not renegotiate every time. Every new icon has to come back matching that reference, and if it drifts, I regenerate it rather than keep it because it's close enough. This mirrors what people who build flat icon sets already know: consistency comes from shared constraints you refuse to break, an identical grid, a fixed stroke, a template you do not touch. With AI, the template is your style reference and your willingness to throw away everything that misses it.
Flat, head-on framing helped more than I expected. Models love to sneak in a little perspective, and those tiny distortions read as inconsistency when you scroll a grid of them. Forcing every icon to sit at the same flat angle made the set feel like one family instantly.
Pick your look once and defend it
The single biggest lever for a cohesive set is the light, and with AI that means a reference you enforce ruthlessly. I settled on a soft key from roughly the upper left, a gentle fill so the shadows don't go muddy, and one subtle highlight for a bit of life. That's it. No dramatic rim lights, no per-icon mood.
The temptation, especially on a hero icon you're excited about, is to let the model do something special because it deserves it. Resist that. The moment one icon has its own lighting, it stops belonging to the set. I keep one reference image that defines where the light comes from, and anything that doesn't match it doesn't ship. If you can't tell me which corner your light comes from without checking, your set is already drifting.
There's a related lesson about restraint. A few soft highlights and one confident color almost always beat a busy, over-detailed render, especially once the icon shrinks to a small size in a real interface. Detail you can't see at 32 pixels isn't craftsmanship, and with a generator it's not even effort. It's just noise the model added that you then have to clean back out.
Naming and search are half the product
Here's the thing nobody warns you about: a library of 2,600 icons is worthless if people can't find the one they need. I underestimated this badly. For a while I named files however I felt in the moment, and then I couldn't find my own icons. If the person who made the library is searching for three minutes, the customer has already given up.
So I rebuilt around search, and this part is all hands, no model. Every icon carries a clear primary name, a category, and a spread of keywords covering the words real people actually type. A lightbulb isn't just "lightbulb," it's idea, innovation, energy, tip, insight. An alien is sci-fi, space, character, game. The goal is that whatever word is in your head when you need the icon, the library meets you there. I now spend nearly as long on an icon's metadata as on the icon itself, and it's the least glamorous work in the whole thing. It's also what makes the library feel like it has 2,600 icons instead of a haystack with 2,600 needles.
A practical tip if you're building anything similar: write the keywords as if you're the frustrated user, not the proud maker. You know it's an "asteroid mining drone." They're typing "robot" and "space" and hoping.
Soft 3D is a trend with a trap
Soft 3D icons are having a real moment. They feel warm and approachable, they work beautifully for creative tools, lifestyle apps, and increasingly for brands that want to feel human without feeling cheap. That popularity is part of why I leaned into the style, and, honestly, it's a style current AI models are unusually good at, which is the other half of why. But there's a trap inside the trend.
The playful, glossy feel doesn't fit everywhere. Drop a bouncy 3D icon into a serious financial dashboard or a dense enterprise tool and it can read as toy-like, out of place, undermining the very trust the product is trying to build. I learned to design the library so the icons hold up when used sparingly and with intent, as accents that draw the eye to the one thing that matters on a screen, not as confetti scattered across every button. One well-placed 3D icon does more work than ten fighting for attention.
My rule of thumb now: if you wouldn't put a photograph there, think twice before putting a soft 3D icon there. Both carry weight. Both slow the eye down. That's a feature when you use it on purpose and a bug when you don't.
Ship in a format people can actually use
An icon that looks stunning in a preview but arrives as a giant file with a hard white box around it is useless to a designer, and a raw generation is almost always exactly that. Two decisions mattered here. First, transparent backgrounds, always, so an icon drops onto any surface without a rectangle fighting the layout. That means cleaning the background out of every output, not hoping the model left one behind. Second, sensible, web-ready file sizes, because nobody wants a 3D icon that costs them a second of load time.
I also learned to provide more than one size. A large display render for hero moments and a small optimized version for interface use are different jobs, and pretending one file does both leaves you with something either too heavy for the small case or too soft for the big one. Serving both respects the person's actual workflow instead of my convenience.
The library is never finished, and that's fine
For a long time I treated the library as a thing I'd one day complete. Reach some number, declare victory, move on. That framing was a mistake, and it's a tempting one when generating a new icon is cheap. The discipline isn't producing more; anyone with a model can flood a folder. It's deciding what earns a place and what gets cut. A living library grows toward its users. The number on the homepage matters far less than whether the next person who searches finds their word waiting for them.
On being upfront about the AI part
I went back and forth on how much to say about how these are made. The quiet option was to let the soft renders speak and never mention the model. I don't think that holds up. People can usually tell, and the ones who can't deserve to know anyway. Saying it plainly costs me nothing real, and it's the difference between a library you trust and one you feel slightly tricked by later. The AI generates the raw material. The judgment is where the actual product lives: the locked style, the ruthless cutting, the metadata, the restraint, the formats. None of that is automatic.
What I'd tell my earlier self is simple. Lock the system before you scale it. Spend real time on the words, not just the pixels. Let restraint be your default. Say what the tool is instead of hiding it. And treat the whole thing as a product you tend, not a collection you finish. Two thousand six hundred icons in, that's the lesson that keeps paying off.