Reproduce “DNA: Dual-stage Native Attribution for Generated Image Source Tracing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chao Wang, Kejiang Chen, Zijin Yang, Yaofei Wang, Yuang Qi, Weiming Zhang, Nenghai Yu
DNA (Dual-stage Native Attribution) presents a training-free framework for tracing AI-generated images back to their specific source models, including highly similar "within-family variants." It achieves an 89.11% end-to-end attribution accuracy on the novel DNA-30K benchmark by leveraging distinct attribution signals from a model's VAE for coarse-grained family-level screening and its generative backbone for fine-grained model-level identification.
Key insights
- Attribution signals within latent generative models are hierarchically stratified, with VAEs capturing family-shared information and generative backbones (e.g., U-Net, velocity predictor) encoding variant-specific differences.
- A coarse-to-fine attribution pipeline, leveraging this signal hierarchy, allows for efficient open-set family-level screening before executing precise closed-set within-family variant discrimination.
- Native prediction consistency, measured at carefully selected discriminative timesteps under semantic conditioning, provides robust and intrinsic signals for distinguishing highly similar generative model backbones without requiring additional neural network training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.