Transformer method improves detection of edited image parts
DINO-FIL: Multi-Layer Transformer Fusion for Generative Inpainting Forgery Localization
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Finding fake parts in images is tricky because the changes can be tiny and blend in well. The authors created DINO-FIL, a method that uses multiple layers from a powerful AI model to spot these subtle edits more accurately. By combining details from different steps inside the model, they catch small clues that other methods miss. They tested it on many types of edited images and it worked better than older tools while not falsely accusing real images.
What this means in practice
- •For digital forensics teams: Detect subtle image edits to locate forged regions in photos during investigations.
- •For photo editing software developers: Integrate advanced forgery detection to warn users about possible image inpainting manipulations.$Commercial implications: This technique enables superior forgery alerts that can be sold as a feature to software makers for digital image authentication.
Authors
Simon Freitter, David Fischinger, Silvia Poletti, Marcel Hasenbalg, Martin Boyer
Abstract
We propose DINO-FIL, a transformer-based framework for pixel-level localization of generative image forgeries. The method leverages multi-layer representations from a pretrained DINOv3 ViT-H+/16 encoder combined with a lightweight fusion decoder to capture both semantic structure and fine-grained texture inconsistencies introduced by image editing. By aggregating intermediate transformer features across depths, the model improves sensitivity to subtle, spatially localized artifacts that are often missed by standard single-layer approaches. We evaluate DINO-FIL in a cross-dataset setting, training on B-Free and testing on Beyond the Brush, CocoGlide, TGIF, and SAGI-D. The model achieves strong generalization, outperforming existing baselines on average and on four of five evaluated datasets, while maintaining a high true negative rate on pristine images.