Generative image forgeries are becoming harder to spot, and locating exactly which pixels have been manipulated is a challenging task. A new paper on arXiv introduces DINO-FIL, a transformer-based framework designed for pixel-level localization of such forgeries.

The method uses multi-layer representations from a pretrained DINOv3 ViT-H+/16 encoder. By fusing these layers, the model aims to capture both fine-grained details and broader contextual information that can indicate where an image has been synthetically altered. The abstract mentions a lightweight component, but the full architecture is not detailed in the available text.

Because the source is only an abstract, no experimental results or comparisons with other methods are provided. The paper's contribution is the combination of a strong pretrained vision transformer with multi-layer fusion for forgery localization, a direction that may be useful for digital forensics.