Text-to-image models get worse the more you ask of them — and a new paper proposes a concrete fix.
Researchers introduced Correlation-Weighted Multi-Reward Optimization (CMO), a training framework designed to tackle one of diffusion models' most persistent failure modes: compositional prompts. When a prompt asks for multiple objects, attributes, or relationships at once, models like SD3.5 and FLUX.1-dev routinely drop or misrepresent some of them. CMO addresses this by decomposing multi-concept prompts into groups — objects, attributes, relations — and assigning separate reward signals to each. It then uses correlation-based difficulty estimation to upweight whichever concepts are proving hardest to satisfy, steering optimization toward the most underperformed parts of the prompt.
The practical upshot is measurable: the authors report consistent gains on ConceptMix, GenEval 2, and T2I-CompBench, three benchmarks specifically designed to stress-test compositional following. That matters because compositional failure is one of the main reasons practitioners still reach for manual prompt engineering, inpainting, or multi-step pipelines instead of a single generation pass.
The approach is conceptually similar to curriculum learning — spending more training effort where the model is weakest — applied to reward signals rather than data. Whether these benchmark gains translate to the messier, open-ended prompts real users write is the usual open question. Code is public on GitHub, so that answer should arrive soon enough.