ForensicSAM: ELA-Guided Prompt-Free Forgery Segmentation with Real-Time SAM Adaptation

Abstract—Diffusion-based manipulations and advanced splic-ing techniques continue to challenge conventional CNN-basedimage forgery detectors, especially when precise pixel-level lo-calization is required. To address this problem, we proposeForensicSAM, a prompt-free forgery segmentation frameworkthat adapts the Segment Anything Model using Error LevelAnalysis (ELA)-guided cues. The method is trained and evaluatedon 27,651 images from CASIA v2.0, CoMoFoD, Columbia, andIMSplice. ForensicSAM achieves an IoU of 96.8 % with aninference time of 6.3 ms, outperforming U-Net, DADF, and stan-dard SAM. The proposed approach eliminates manual prompt-ing by generating automatic bounding-box prompts from ELAvariations and supports zero-shot localization across multipleforgery types, including diffusion-generated content. Ablationresults show that ELA guidance, automatic prompt generation,and the ViT-B encoder each contribute to the final performancegain. Overall, ForensicSAM offers an effective balance betweenlocalization accuracy and inference speed for automated imageforgery detection