Image Editing Evaluation
Overview
Image editing evaluation asks whether an image makes the requested change, preserves everything else, and remains visually coherent across tasks from local retouching to multi-turn and structured-content edits. Benchmarks use human ratings, paired references, controlled edits with pixel- or region-level ground truth, and specialized tests for text, geometry, identity, reasoning, and physical plausibility. Automatic methods range from image-similarity and OCR measures to vision-language model judges, interpretable rubrics, reward models, and localized defect scoring, often checked against human judgments. A central challenge is that holistic or automated scores can miss subtle unintended changes and task-specific errors, while exact metrics do not transfer across settings; evaluation must balance fine-grained diagnosis with scalable human alignment.
Latest papers
92 papers16 papers in the last four weeks, up 300% on the four weeks before. 0.1% of all new papers.