What it is
Unified event-to-video diffusion pipeline for long-horizon reconstruction, prediction, and interpolation from sparse event streams. Turn sparse event voxel streams into stable videos for reconstruction, future prediction, and frame interpolation with one inference and training pipeline.
Key results
- Supports 3 tasks: reconstruction, prediction, and interpolation
- Released preprocessed data for BS-ERGB, ECD, MVSEC, and HQF
- Provides released LoRA weights and full generated eval outputs
Shows that a single diffusion-based pipeline can handle long-horizon event-to-video generation across multiple tasks.