What it is
Maps final video-model VAE latents directly to dynamic 3D scenes, skipping RGB reconstruction. One trained checkpoint transfers across multiple compatible video generators. Instead of decode-video-then-reconstruct or retraining each generator for geometry, it treats shared final VAE latents as the interface to 4D prediction and learns one latent-to-4D adapter plus refinement stack that works across compatible DiTs.
Key results
- I4D-200: 61.60 DINO-F1, +5.81 over matched Wan2.2-I2V-A14B+4RC
- Text4D-200: 57.09 DINO-F1, +2.88 over matched Wan2.1-1.3B+4RC
- Text4D-200: 57.01 DINO-F1, +3.45 over matched Wan2.1-14B+4RC
Shows shared VAE latents can serve as a reusable geometry interface across compatible video generators without generator-specific 4D retraining.