StableI2I: Spotting Unintended Changes in Image-to-Image Transition

Researchers introduced StableI2I, a reference-free framework for image-to-image evaluation that measures content preservation and pre/post consistency.

It checks whether outputs keep the input’s semantics and spatial structure, helping detect unintended changes beyond style or visual quality.

The team also released StableI2I-Bench, a benchmark for testing how accurately MLLMs judge these fidelity and consistency cases.

The work fills a key evaluation gap and offers a more reliable way to compare models against human perception.