Most text-to-3D generators are focused on single mesh generation, which can be difficult to edit post-generation. While less suitable for character or creature assets, the ThreeJS Generator is particularly useful for creating architectural or prop assets, with modular, animated, or procedural elements. As an open source tool not gated by paid subscriptions or opaque credit systems, anyone can generate editable 3D assets, with generation costs paid directly to providers using their own API keys.
What real task does your product handle with GPT-6 Astra?
Maker
GPT-6 Astra makes 3D asset generation finally genuinely viable for actual development workflows, and this product makes that open source, with API costs going directly from the user to the API provider, with no opaque subscription or credit system business models in-between.
Report
Maker
📌
In January, I built an earlier iteration of this codebase locally with Cursor, and found the results promising (particularly with regard to the modular/procedural elements), but the models clearly weren't quite there yet with their spatial reasoning. Most of the generations would come out at least partially broken, or with strange placement, so I decided to table it for the time being.
After seeing OpenAI's 3D development demo for GPT-6 Astra upon release, I decided to see how much a simple model bump would do in the existing repo -- and was astonished by the results. Rather than keeping it to myself and sitting on the codebase while trying to figure out monetization as a for-profit product, I decided to spend the last week shaping it up with Codex and preparing it for an open source release.
With this first release, only OpenAI and Anthropic are supported as the model providers, with a bring-your-own-API-key structure for provider-direct billing. For the best results, I'd recommend GPT-6 Astra. In my own tests (and your mileage may vary -- not a guarantee of costs, etc), Astra has been running about ~1-2 minutes and ~$0.40 per generative request. Fable 5.1 tends to run longer, with some generations running over five minutes, and has been more costly, generally upward of a dollar per request. 3D model quality also currently seems to favor GPT in my tests. That said, provider performance does seem to be a game of leap frog, and the app should scale organically as new models are released with further improved 3D capabilities, making it an interesting visual benchmark tool as well. Local models are not currently supported for generation, but that's something I'll explore adding in a future release.
So far I've found the generations to be -- at the very least -- very workable starting points for 3D assets, and have included two demos in the repo that showcase the raw procedural ThreeJS code in use. I'll be using this in my own AI-assisted game development (and have already started, with prop assets created for the Unity-based game I've been working on in Codex), and I'm extremely interested to see what assets others might generate with it, for their own workflows.
Report
Maker
UPDATE: You can now optionally log into your OpenAI account and tap your Codex subscription usage for 3D model generation instead of paying API costs. That should make the repo even more viable for democratizing 3D model generation -- from the Codex side panel:
UPDATE: You can now optionally log into your OpenAI account and tap your Codex subscription usage for 3D model generation instead of paying API costs. That should make the repo even more viable for democratizing 3D model generation -- from the Codex side panel: