Text2Gaussian is an open-source pipeline that converts text descriptions into 3D Gaussian Splat models. Type a prompt like "a vintage brass teapot" and get a fully-formed 3D asset in approximately 17 seconds—output as a PLY file with 200,000+ vertices.
Overview
Pipeline
How It Works
The pipeline operates through three sequential stages:
- Text to Image — Google's Gemini AI generates images from text descriptions
- Image to 3D — Meta's SAM 3D Objects model processes the generated images
- Output — Results are saved as PLY files compatible with standard 3D viewers
Architecture
- stage0/ — Batch processing pipeline for generating multiple assets
- stage1/ — Text-to-image generation via Gemini integration
- runpod-server/ — GPU-accelerated 3D conversion API running on A100
- viewer/ — Web-based 3D visualization interface