Back to Pixels: Lifting Latent Video Diffusion into Pixel Space

Researchers have introduced PixVid, a method that lifts pretrained latent video models to directly generate every frame in pixel space using a convolutional head instead of a VAE decoder. By predicting clean video rather than velocity and applying a larger noise shift, PixVid achieves high quality competitive with latent baselines. Finetuned from Wan2.1-14B, the model supports higher resolution generation, improved fast-motion depth editing, and per-frame action responsiveness. Open article →