Spaces:
Running on Zero
Running on Zero
| title: DIRECT 3D-Aware Object Insertion | |
| emoji: 🪑 | |
| colorFrom: gray | |
| colorTo: red | |
| sdk: gradio | |
| sdk_version: 5.50.0 | |
| app_file: app.py | |
| short_description: 3D-aware object insertion with the DIRECT model (ICML 2026) | |
| python_version: "3.10" | |
| startup_duration_timeout: 1h | |
| pinned: false | |
| # DIRECT: Direct 3D-Aware Object Insertion via Decomposed Visual Proxies | |
| Insert a reference object into a background scene with realistic, harmonized | |
| results using the [DIRECT](https://huggingface.co/superGong/DIRECT) model | |
| (ICML 2026), a FLUX.1-Fill-dev network guided by a decomposed visual proxy. | |
| Upload a background image and an object image (its background is removed | |
| automatically), choose where and how large to place the object, then click | |
| **Insert**. | |
| ## Note on scope | |
| The full paper uses an interactive 3D viewer (TRELLIS + Viser) so users can pose | |
| a reconstructed 3D proxy of the object. That live 3D websocket viewer cannot run | |
| inside a single-port Hugging Face Space, so this demo drives the same DIRECT | |
| model with a simpler **2D placement** (position + scale) as its geometric | |
| guidance. The insertion / harmonization is performed by the real DIRECT weights. | |
| ## Links | |
| - Paper: https://arxiv.org/abs/2606.06601 | |
| - Project page: https://gong1130.github.io/DIRECT/ | |
| - Code: https://github.com/Gong1130/DIRECT | |
| - Model weights: https://huggingface.co/superGong/DIRECT | |