Context Unrolling in Omni Models

Ceyuan Yang, Zhijie Lin, Yang Zhao +16

We present Omni, a unified multimodal model natively trained on diverse modalities, including text, images, videos, 3D geometry, and hidden representations. We find that such train…