JUICE: Unified Multimodal Modeling in Pretrained Representation Spaces. Pretrained (Base) and SFT models, and the frozen image decoders they use.