DINO-FIL: Multi-Layer Transformer Fusion for Generative Inpainting Forgery Localization

Key points

• 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.1
• 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.1
• We evaluate DINO-FIL in a cross-dataset setting, training on B-Free and testing on Beyond the Brush, CocoGlide, TGIF, and SAGI-D.1
• 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.1

Sources (1)

• [1]DINO-FIL: Multi-Layer Transformer Fusion for Generative Inpainting Forgery Localization
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 12:36 PM

  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.

Extractive summary: sentences quoted from the sources.