Researchers introduce PixVid, a method that lifts pretrained latent video diffusion models into pixel-space generation by replacing the VAE decoder with a convolutional head rendering framewise outputs directly in pixels. Finetuned from Wan2.1-14B using approximately 1.3 million training clips, PixVid maintains competitive quality with the latent baseline while resolving key limitations of latent-space compression. The approach predicts clean video frames rather than velocity and employs a larger noise shift, which the authors identify as critical for quality. Operating in pixel space enables generation at twice the spatial resolution with equivalent per-frame token counts, since latent compression typically quadruples token requirements when doubling resolution. Framewise processing also allows conditions to apply to individual frames rather than frame groups, and each response occurs at the next frame rather than waiting for multi-frame block completion. Demonstrated capabilities include depth-conditioned video editing that remains aligned under fast motion and interactive generation responsive to each action at the subsequent frame. The method retains the original diffusion transformer architecture while modifying only the decoding pathway, offering a simple yet effective path to high-resolution, controllable, and low-latency video synthesis.

Read original