ReMoA Framework for Recursive Reasoning in Multimodal Diffusion

The ReMoA framework introduces recursive modular adaptation to diffusion models via joint attention layers. By iteratively refining visual tokens through sparse neural module selection, it improves complex text-following capabilities in multimodal generation.