UniRect-CoT: Enhancing Generation in Unified Multimodal Models via Reflective Rectification with Inherent Understanding
UniRect-CoT addresses Unified Multimodal Models' weakness where their understanding is strong but generation is weak. It introduces a training-free 'Thinking-While-Drawing' method that uses the model's existing knowledge to reflect on intermediate image steps. This continuous self-correction process rectifies errors, unlocking the model's full generative potential for complex tasks.