[Feature Request] Spatial Temporal Consistency

[tomhalpin8]

1. Video Support & Temporal Stability
   Extend the model to accept video input rather than just static images. By utilizing information between frames (inter-frame data), the generation of splats can become more stable and consistent over time, reducing jitter or flickering.

2. Improved Background Depth & Rigidity
   Enhance the representation of rigid objects in the background when the camera is moving. The model should better distinguish between the moving camera and static background elements to maintain accurate depth.

3. Bidirectional Training for Disocclusion
   Implement forward and backward training passes. If a subject moves and reveals (disoccludes) the background, the model should use future frames to backfill that visual information, ensuring the background remains consistent throughout the entire specific shot.

4. File Compression
   Introduce a compression standard for the output. Currently, a single second of footage results in ~1.5 GB files, which makes video generation impractical for real-world usage.