GenText-Forensics: Challenge on Explainable Forensics and Adversarial Generation for Text-Centric Images
Abstract
AI-generated content has lowered the cost of forging text-centric images, threatening financial, legal, and information systems. Building on DocShield and its RealText-V1 benchmark, we develop RealText-V2 and launch the GenText-Forensics Challenge at ACM Multimedia 2026 to advance reasoning-based text-centric image forensics. The challenge formulates forgery analysis report generation as a unified task that integrates binary forgery detection, pixel-level tampering localization, and evidence-grounded explanation in a single structured report. RealText-V2 scales this formulation to broader languages, application domains, and manipulation types. Submissions are evaluated by detection F1, localization mF1, explanation BERTScore, and LLM-judged report quality. Our analysis shows that leading systems consistently adopt modular expert pipelines, localization remains the primary performance driver, and evidence conditioning is central to reliable MLLM reasoning. To foster open research on AI document safety, we publicly release RealText-V1, RealText-V2, and multi-scale DocShield model checkpoints through a unified collection.