| # 🌀 Spatial Diffusion | |
| **Spatial Diffusion** is a generative model for synthesizing **spatial panoramas** based on a **cubemap representation**. By generating six orthogonal cube faces (front, back, left, right, top, bottom), the model constructs a complete and spatially consistent 360° view of a scene. This cubemap-based approach ensures geometric coherence and enables immersive scene generation for various downstream applications. | |
| ## 🌐 Model Highlights | |
| - **Cubemap Representation** | |
| Generates six cube faces to represent the entire spherical environment, maintaining consistent spatial alignment. | |
| - **Diffusion-Based Generation** | |
| Uses a diffusion process to progressively refine spatial details and structure, producing high-quality and coherent outputs. | |
| - **360° View Synthesis** | |
| Capable of producing panoramas suitable for virtual reality, robotics, and simulation environments. | |
| ## 🚀 Intended Applications | |
| - Virtual Reality (VR) scene generation | |
| - Environmental simulation and reconstruction | |
| - Robotics & autonomous navigation (spatial awareness) | |
| ## ⚠️ Limitations | |
| - Performance may drop in scenes with non-Euclidean geometry or extreme occlusions. | |
| - Post-processing may be required for equirectangular projection if not viewed via cubemap renderers. | |
| - May not generalize well outside the distribution of the training dataset. | |
| ## 📄 Citation | |
| If you use this model in your research or application, please cite: | |
| Spatial Diffusion: Cubemap-Based Generation of Spatial Panoramas, [Ziming He], 2025. | |