GPT-6 Astra builds a Three.js character, then Tripo turns it into a usable 3D model
The author first used Astra to build a character prototype in Three.js, then compared the mesh quality and editability of a dedicated 3D model generator.
I wanted to try something a little more personal with this one. Most of you have seen the Token Gremlin in one form or another by now, so I decided to take that character and see how far I could push it through an AI-assisted 3D workflow. I started with GPT-6 Astra. I gave Astra the original character reference and had it build a first interpretation of the Gremlin inside Three.js. It was already enough to turn the idea into something interactive and give me a rough sense of how the character could exist in a 3D environment. But that also made me curious about something else: what happens when you hand the actual 3D generation work to a model that is specialized for it? So I took the same character into Tripo. That is where the experiment became much more interesting. Instead of stopping at something that simply looks 3D from the outside, Tripo gave me an actual model I could inspect, rotate, texture, rig and animate. I could go from the original visual reference to a usable 3D asset without manually modeling the character from scratch. And this is probably the part that matters most to me as someone who builds games, Three.js experiments and weird interactive projects all the time. Generating a pretty object is one thing. Generating geometry that can actually survive the next stages of a production pipeline is much more useful. Tripo’s Smart Mesh P2.0 is especially interesting here because it is focused on the mesh itself, not just the final render. It can generate native quad topology with cleaner edge flow and controllable polygon budgets, so the geometry is much closer to something you can continue working with instead of treating the AI output as the final destination. For game and real-time workflows, that distinction matters a lot. Clean quads deform more predictably when rigged and animated, subdivision behaves better, and you spend less time fighting a chaotic mesh before the asset can actually be used. Smart Mesh P2.0 also lets you control topology density, with more geometry where the model needs detail and cleaner surfaces where it does not. That is the direction I find much more interesting for generative 3D. Not “AI made a cool picture of a model.” More like: idea → generation → usable mesh → texture → rig → animation → actual project. And this experiment also reinforced something I keep noticing with AI tools: I get much better results when I stop expecting one model to do absolutely everything. Astra was useful for helping me shape the initial interactive workflow and translating the idea into a Three.js experiment. Tripo was much more interesting as the specialized 3D layer, where the character could become an actual asset and move further down the production pipeline. Let each tool do the thing it is actually good at. The final Gremlin ended up looking considerably better than I expected, and seeing my little creature go from a flat reference to an animated 3D character was honestly pretty damn fun. If you work with games, 3D, Three.js, Blender or just like experimenting with generative tools, Tripo is worth playing with.