NVIDIA open-sources PixelUMM, unifying video and image understanding and generation in pixel space.

NVIDIA open-sources PixelUMM for unified video and image understanding and generation, adopting pixel space representation instead of VAE and ViT, placing understanding and generation tasks in the same context, reducing token consumption and improving efficiency. The model supports video and image inputs, processing visual tasks with a single token to achieve more compact multimodal interaction.