ByteDance Seed removes the VAE bottleneck from unified multimodal models

Their technique, Representation Forcing, lets decoders predict visual representations before pixels so generation and understanding share one end-to-end space.