Summary
PixelUMM is a unified model for understanding and generating images and video directly in pixel space. A single decoder-only transformer processes image patches and video patches without a separate vision encoder or VAE.
PixelUMM is a unified model for understanding and generating images and video directly in pixel space. A single decoder-only transformer processes image patches and video patches without a separate vision encoder or VAE.
Researchers can study a shared visual model that reads and writes pixels while keeping the generation and understanding paths within one architecture. Core capabilities include Image understanding, Video understanding, Image generation, and Video generation.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Raw-pixel modeling, No separate vision encoder, No VAE image bottleneck, and Public implementation.
Get more likes & reach the top of search results by adding this button on your site!


Get top updates in AI to your inbox every weekend. It's free!