[2604.03400] Banana100: Breaking NR-IQA Metrics by 100 Iterative Image Replications with Nano Banana Pro

Computer Science > Computer Vision and Pattern Recognition arXiv:2604.03400 (cs) [Submitted on 3 Apr 2026] Title:Banana100: Breaking NR-IQA Metrics by 100 Iterative Image Replications with Nano Banana Pro Authors:Kenan Tang, Praveen Arunshankar, Andong Hua, Anthony Yang, Yao Qin View a PDF of the paper titled Banana100: Breaking NR-IQA Metrics by 100 Iterative Image Replications with Nano Banana Pro, by Kenan Tang and 4 other authors View PDF HTML (experimental) Abstract:The multi-step, iterative image editing capabilities of multi-modal agentic systems have transformed digital content creation. Although latest image editing models faithfully follow instructions and generate high-quality images in single-turn edits, we identify a critical weakness in multi-turn editing, which is the iterative degradation of image quality. As images are repeatedly edited, minor artifacts accumulate, rapidly leading to a severe accumulation of visible noise and a failure to follow simple editing instructions. To systematically study these failures, we introduce Banana100, a comprehensive dataset of 28,000 degraded images generated through 100 iterative editing steps, including diverse textures and image content. Alarmingly, image quality evaluators fail to detect the degradation. Among 21 popular no-reference image quality assessment (NR-IQA) metrics, none of them consistently assign lower scores to heavily degraded images than to clean ones. The dual failures of generators and evaluators ma...