README.md

metadata

title: TEXTure CPU Lite
emoji: 🎨
colorFrom: green
colorTo: pink
sdk: gradio
sdk_version: 6.2.0
app_file: app.py
pinned: false
license: mit

TEXTure CPU Lite - Text-Guided 3D Texturing

Generate proper UV texture maps for 3D meshes using text prompts. Uses SD-2-Depth (same as original TEXTure paper) for depth-conditioned generation and xatlas for UV unwrapping.

Features

• Proper UV Texture Output: Creates UV atlas that can be applied to the mesh
• SD-2-Depth: Native depth conditioning (same model as original TEXTure)
• OBJ Export: Outputs OBJ + MTL + texture PNG ready to use
• xatlas UV Unwrapping: Automatic UV coordinate generation

How it works

1. Upload a 3D mesh (.obj, .stl, .ply, .glb)
2. Enter a text prompt describing the desired texture
3. xatlas generates UV coordinates
4. Depth is rendered from multiple viewpoints
5. SD-2-Depth generates textures conditioned on depth
6. Textures are projected back to UV space
7. Download textured mesh (OBJ + MTL + PNG)

Models Used

Component	Model	Size
SD-2-Depth	radames/stable-diffusion-2-depth-img2img	~5GB

Same architecture as original TEXTure paper (SD-2-Depth) - has native depth conditioning built-in. Uses public community copy since official stabilityai/stable-diffusion-2-depth is gated.

Runtime:

• First run: Downloads 5GB model (cached in `/.cache/huggingface/`)
• CPU uses INT8 quantization via optimum.quanto for 3-5x speedup
• ONNX not supported (SD-2-Depth has 5-channel UNet)

Debug logs: Check build logs for [OK], [ERROR], [WARN] messages if something fails.

Performance

Device	Steps	Time per View	4 Views Total
GPU (CUDA)	20	~2-3 sec	~12 sec
CPU (INT8)	10	~1.5 min	~6 min
CPU (INT8)	20	~3 min	~14 min

Tip: For faster results on CPU, use fewer steps (5-10) or fewer views (2-3).

Local Development

pip install -r requirements.txt
python app.py

Files Structure

├── app.py              # Single-file implementation
├── requirements.txt
├── README.md
└── shapes/
    └── bunny.obj       # Sample mesh

Credits

• TEXTure Paper - Yael Vinker et al.
• ControlNet - Lvmin Zhang
• Stable Diffusion - RunwayML
• xatlas - UV unwrapping