Most Unified Multimodal Models still generate images through a frozen VAE, which means perception and generation are not fully learned in one model. This paper fixes this by making the decoder first predict

→ View original post on X — @askalphaxiv