New CLIP/DINOv2 study reveals optimal layers for AI image attribution

A new paper from researchers Meiling Li, Pietro Bongini, Benedetta Tondi, and Mauro Barni explores how to identify which generator created an AI-generated image without retraining classifiers. The training-free reference-based method works by comparing suspicious images against a set of known references from each generator. The study systematically examines two coupled factors: the representation space (which layer of CLIP or DINOv2 to use) and how references are constructed. Three reference selection methods are tested: arbitrary (random samples from a generator), semantically aligned (matching semantic content of queries), and resynthesis-based (creating references by re-generating images from a source).

Results show attribution accuracy consistently peaks at intermediate layers of both CLIP and DINOv2, before strong semantic abstraction dominates. This suggests source-discriminative cues are most accessible at mid-level representations. However, these intermediate layers are not semantically neutral, making reference selection critical. Semantically constrained references reduce mismatches between query and reference sets, significantly improving accuracy, especially with limited references. Resynthesis-based references are most useful when only a few references per generator are available. When a moderate-sized reference pool is available, semantically aligned references provide the best accuracy-cost trade-off. The findings emphasize that training-free attribution should be understood as a joint optimization of where (representation layer) and how (reference construction) comparisons are made.