zimhe/SpatialDiffusion

🌀 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.