Announcement_22

I’m glad to share our ICML 2026 work StableI2I, a fidelity-oriented evaluation framework for image-to-image generation. Rather than only asking whether an edited/restored image looks good or follows the instruction, StableI2I focuses on what has been unintentionally changed. It jointly considers the input image, output image, and I2I instruction to diagnose content drift across semantic consistency, structural fidelity, and low-level appearance, covering errors such as object addition/removal/replacement, repainting, misalignment, noise, blur, and color cast. We release StableI2I-Bench, together with StableI2I and StableI2I-PLUS models, to support fine-grained I2I fidelity diagnosis and scoring for more faithful and controllable image editing/restoration systems. [Homepage] [GitHub] [StableI2I-Bench] [StableI2I Model] [StableI2I-PLUS] [Paper].