Beyond Pixels: From Video Priors to 4D Worlds
The paper proposes using denoised video latents as a reusable bridge into explicit 4D scene generation.
Its Latent-to-4D method skips RGB reconstruction and aligns video-model latents with a pretrained 4D decoder’s token grid. Trained on roughly 1,000 reconstruction clips, one checkpoint transfers across multiple video diffusion transformers that share the same VAE family. The authors report gains over matched Wan+4RC cascades on Text4D-200 and I4D-200, plus human preference for geometry, temporal stability, and overall quality. Source: HF Daily Papers' note.
Its Latent-to-4D method skips RGB reconstruction and aligns video-model latents with a pretrained 4D decoder’s token grid. Trained on roughly 1,000 reconstruction clips, one checkpoint transfers across multiple video diffusion transformers that share the same VAE family. The authors report gains over matched Wan+4RC cascades on Text4D-200 and I4D-200, plus human preference for geometry, temporal stability, and overall quality. Source: HF Daily Papers' note.
score 5