Today's top video generators work inside a compressed 'latent' space, a kind of shorthand, so they can train faster. But this compression creates three problems: doubling the video's resolution quadruples the number of tokens (the data chunks the model processes); adding conditions like depth maps affects entire blocks of frames instead of individual ones; and each response must wait for a whole block to finish before moving on.
The researchers finetuned a pretrained latent model, Wan2.1-14B, to instead generate every frame directly as pixels. They replaced the part that decodes the latent space with a simple convolutional head, a network that turns the model's internal tokens into pixel colors. Key changes: predicting the clean pixel values rather than the noise, and using a larger noise shift during training.
After training on just 1.3 million clips, PixVid matched the original latent model's quality. It generates at double the resolution with the same per-frame token count, preserves 98% of fast motion under depth-based editing (versus 83%), and cuts control-to-screen latency from 495 ms to 34 ms. This makes pixel-space generation practical for interactive and high-resolution video.