StableI2I Framework Detects Unwanted Changes in Image-to-Image Transitions

Researchers have introduced StableI2I, a unified evaluation framework that measures content fidelity and pre-post consistency in image-to-image (I2I) tasks without reference images. Unlike existing evaluations that focus on instruction following and perceptual quality, StableI2I explicitly assesses whether output images preserve semantic correspondence and spatial structure. The framework covers a wide range of I2I scenarios, including image editing and restoration. Alongside the framework, the team built StableI2I-Bench, a benchmark to systematically evaluate multimodal large language models (MLLMs) on fidelity and consistency assessment. Experimental results show StableI2I provides accurate, fine-grained, and interpretable evaluations with strong correlations to human judgment. The work addresses a critical gap in I2I evaluation by ensuring that generated images maintain intended content and structure.