This is very neat from ‘MMaDA-Parallel: Multimodal Large Diffusion Language Models for Thinking-Aware Editing and Generation’