Spaces:
Running on Zero
Running on Zero
| title: VEGA-3D Spatial Reasoning | |
| emoji: 🧭 | |
| colorFrom: red | |
| colorTo: gray | |
| sdk: gradio | |
| sdk_version: 5.49.1 | |
| app_file: app.py | |
| short_description: 3D spatial reasoning VLM with video-diffusion 3D priors | |
| python_version: "3.10" | |
| startup_duration_timeout: 1h | |
| pinned: false | |
| # VEGA-3D · Spatial Reasoning | |
| Interactive demo of [**VEGA-3D**](https://huggingface.co/H-EmbodVis/VEGA-3D-Spatial-Reasoning), | |
| a spatial-reasoning multimodal LLM from the paper | |
| [*Generation Models Know Space: Unleashing Implicit 3D Priors for Scene | |
| Understanding*](https://huggingface.co/papers/2603.19235) (ECCV 2026). | |
| VEGA-3D augments a Qwen2.5-VL backbone with **implicit 3D priors** extracted | |
| from a frozen **Wan2.1-T2V-1.3B** video diffusion model. Intermediate | |
| spatiotemporal features from the diffusion model are fused into the VLM via | |
| token-level gated fusion, giving it stronger geometric and spatial | |
| understanding of indoor scenes. | |
| Upload a scene image (or a short scene video) and ask a spatial-reasoning | |
| question. | |