AesCase: Towards Interpretable Image Aesthetic Assessment via Badcase Discovery

Most existing image aesthetic assessment methods predict an overall score or preference, yet fail to answer where the problem lies, why it occurs, and how to improve it. To repair a flawed image, a model must first generate concrete repair suggestions, which in turn depends on precisely discovering and localizing its aesthetic flaws. To address this issue, we propose AesCase, a framework for aesthetic badcase discovery: 1) we propose a score-guided five-tier construction method that simulates the human badcase distribution to balance annotation efficiency and quality; building on it, we construct the AesCase-20k dataset and the AesCase-bench benchmark; 2) we train a baseline model AesCase-8B on mixed flaw-discovery and scoring data, jointly learning flaw discovery and score prediction, and generalize flaw localization into repair suggestions (suggestion). Experiments show that AesCase-8B achieves the best flaw-discovery F1 on AesCase-bench (0.726), while producing almost no false positives on good images (0.02 vs. GPT-5.5's 4.05), and maintains competitive scoring performance on six IAA/IQA benchmarks. Furthermore, this repair-oriented, interpretable feedback closes the loop from “where is the problem” to “how to fix it”, making AesCase a plug-and-play model for image editing and aesthetic enhancement.