Spaces:
Running on Zero
Running on Zero
| title: FlexPart | |
| emoji: 🧩 | |
| colorFrom: red | |
| colorTo: blue | |
| sdk: gradio | |
| sdk_version: 6.30.0 | |
| app_file: app.py | |
| short_description: Image-to-3D part-level generation with variable granularity | |
| python_version: "3.12" | |
| startup_duration_timeout: 30m | |
| # 🧩 FlexPart — Image-to-3D Part Generation | |
| **A Unified Framework for Image-to-3D Part Generation via Variable Granularity** | |
| ([paper](https://huggingface.co/papers/2610.31671) · [code](https://github.com/Gucci233/FlexPart) · [weights](https://huggingface.co/gucci233/FlexPart)) | |
| Upload an object image, click once on each part you want, and FlexPart generates a | |
| 3D mesh for every clicked part — composed into a single colored GLB, with each part | |
| also downloadable individually. | |
| ## How it works | |
| 1. **Upload** an object image (or pick a pre-annotated example). The background is | |
| removed automatically with [BriaRMBG-1.4](https://huggingface.co/briaai/RMBG-1.4). | |
| 2. **Click** on the image — every click marks one part (up to 32). The point is used | |
| as a geometric prompt that steers part-level diffusion. | |
| 3. **Generate** — FlexPart runs part-level 3D rectified-flow diffusion | |
| ([TripoSG](https://huggingface.co/VAST-AI/TripoSG) VAE + DINOv2 image conditioning) | |
| and extracts one mesh per part via hierarchical octree marching cubes, then | |
| composes the parts with a per-part color palette. | |
| ## Credits | |
| - Inference code adapted from the official [Gucci233/FlexPart](https://github.com/Gucci233/FlexPart) | |
| repository (MIT license). Example images are the authors' own demo assets. | |