UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation
UFO tests whether generated images satisfy text and visual conditions at the same time, not one condition at a time.
The paper says current evaluators often score each input condition separately, which misses the joint-alignment target of multi-modal image generation. UFO breaks alignment into fine-grained Atomic Evaluation Units, groups them by modality relevance, and verifies them with general or dedicated function calls. The authors report the framework has the highest correlation with human preferences, with an average improvement of 15.25%. They also introduce UFO-Bench for testing customization models under mixed text and visual-condition interactions. Source: HF Daily Papers' note
The paper says current evaluators often score each input condition separately, which misses the joint-alignment target of multi-modal image generation. UFO breaks alignment into fine-grained Atomic Evaluation Units, groups them by modality relevance, and verifies them with general or dedicated function calls. The authors report the framework has the highest correlation with human preferences, with an average improvement of 15.25%. They also introduce UFO-Bench for testing customization models under mixed text and visual-condition interactions. Source: HF Daily Papers' note
score 5