Thinking with Generated Images

Thinking with Generated Images is a novel research paradigm that enables large multimodal models (LMMs) to natively reason across text and vision by spontaneously generating intermediate visual thinking steps. It introduces two complementary mechanisms: (1) vision generation with intermediate visual subgoals, where complex visual tasks are decomposed into manageable components generated and integrated progressively; and (2) vision generation with self-critique, where models generate an initial visual hypothesis, analyze its shortcomings through textual reasoning, and produce refined outputs. The approach achieves up to 50% relative improvement on complex multi-object scenarios in vision generation benchmarks. An open-source suite is released to support further research and applications in fields such as biochemistry, architecture, forensics, and sports strategy.