PixVid retrofits Wan2.1-14B to render video in pixel space

"Frontier video generation models work in a VAE latent space that compresses video in both space and time," the PixVid paper begins. The compression made training tractable, but the authors argue it is now the bottleneck: "doubling resolution quadruples token count," conditions "apply to groups of frames rather than individual ones," and each response "must wait for the rest of its multi-frame block."

Their method, finetuned from Wan2.1-14B, keeps the diffusion transformer but replaces the VAE decoder with a convolutional head that renders directly to pixels. "We find that predicting clean video rather than velocity, together with a larger noise shift, is key to quality," the authors write. After roughly 1.3M training clips they report the pixel-space model is "competitive with the latent baseline within ~1.3M training clips." Working framewise, it generates at "2× higher resolution with the same per-frame token count," keeps depth-conditioned editing aligned under fast motion, and responds to each action at the next frame.

The abstract publishes no numerical benchmark scores against the latent baseline and no per-step inference cost for the new architecture.