XNews

Beijing Academy releases Emu3.5 multimodal world model predicting 'next second'

Beijing Academy releases Emu3.5 multimodal world model predicting 'next second'

Beijing Academy's Emu3.5 is a multimodal world model unifying text, images and video as tokens and framing training as next-state prediction (NSP): predicting the next world state token (visual or linguistic). Emu3.5 focuses on learning spatial relations, temporal object dynamics and multimodal continuity so the model predicts 'the world's next second' across modalities instead of treating image/text/video separately.

Source posts