Back to Pixels: Lifting Latent Video Diffusion into Pixel Space
PixVid generates video at twice the resolution with the same per-frame token count, the researchers report.
The method keeps a pretrained video model’s diffusion transformer and replaces its VAE decoder with a convolutional head that outputs pixels. Fine-tuned from Wan2.1-14B on roughly 1.3 million clips, it is competitive with the latent-space baseline, according to the paper. Predicting clean video rather than velocity, alongside a larger noise shift, proved key to quality. Framewise generation also keeps depth-conditioned editing aligned during fast motion and lets each action affect the next frame.
HF Daily Papers' note
The method keeps a pretrained video model’s diffusion transformer and replaces its VAE decoder with a convolutional head that outputs pixels. Fine-tuned from Wan2.1-14B on roughly 1.3 million clips, it is competitive with the latent-space baseline, according to the paper. Predicting clean video rather than velocity, alongside a larger noise shift, proved key to quality. Framewise generation also keeps depth-conditioned editing aligned during fast motion and lets each action affect the next frame.
HF Daily Papers' note
score 5