November 2024. Built at the BeWater Drinkathon (T3V & BitcoinOS track) in Bangkok over a weekend. Won.
the problem
Creating 3D assets is slow and gatekept by skill. If you're a game developer, a VR designer, or just someone who needs a 3D model of a specific object, your options are: learn Blender properly (weeks), hire someone (expensive), or download something generic from a library (usually not quite right).
The whole pipeline — concept to model — is both time-consuming and skill-dependent. That's the bottleneck.
what we built
A text-to-3D generator that takes a natural language prompt and spits out a usable 3D model. No 3D modelling experience required.
Stack:
- OpenAI Shap-E — the AI backbone. Shap-E is a conditional generative model trained to produce 3D assets from text. It generates both a mesh and a NeRF representation, so you get something you can actually use.
- Gradio — deployed the model as a web app instantly. You type a prompt, hit submit, get a model back. Accessible via browser, no install needed.
- Shap-E Diffuser — post-processing layer that converts the generated mesh and GIF output into a GLB file, which is the standard format for Blender, Three.js, and most 3D workflows.
The demo ran live at ia264745.gradio.live — a simple interface: enter a prompt, clear, submit. The output was a viewable 3D model ready to pull into Blender.
demo
why it mattered
The unlock here wasn't just "AI makes 3D models." It was the full pipeline to a usable file format. GLB is broadly compatible — Blender picks it up natively, it works in Three.js for web rendering, Unity and Unreal can import it. By going text → Shap-E → GLB, we closed the gap from "cool demo" to "actually useful for a 3D workflow."
At scale, this changes who can generate 3D content. You don't need a modeller — you need someone who can describe what they want.
stack
Python, OpenAI Shap-E, Gradio, Shap-E Diffuser (HuggingFace), GLB export
reflection
Shap-E was still early — the models are recognisable but not photorealistic. That's a function of where generative 3D was in 2024, not a flaw in the approach. The pipeline we built is the right one; the underlying model quality will keep improving.
Won the hackathon. Bangkok was a good venue for it.