JVT implements pixel-space video diffusion with a large-patch DiT-style Transformer, without requiring a pretrained autoencoder, auxiliary refiner, cascaded stages, or U-Net skip connections. Experiments show that JVT achieves generation quality competitive with latent diffusion baselines in both class-conditional and unconditional settings; for camera-conditioned RGB-disparity generation, it improves camera-following effects and can directly output scene geometry.
Source:Hugging Face 社区论文精选 · Read original
Leave a Reply