Researchers Simplify Multimodal AI Model By Removing AdaLN

Our simple rule: remove every part that seems to be removable. We starts with pixel space, the standard T5-L encoder, and a simple multimodal MM-JiT backbone with x-prediction.