New Framework Fingerprints Text-to-Image Diffusion Models via Collapsed Generation

A recent study published on arXiv (2608.11732) presents a non-invasive framework for fingerprinting proprietary text-to-image diffusion models. This approach utilizes 'collapsed generation,' a phenomenon where specific input conditions yield remarkably consistent images across various stochastic seeds, reflecting the model's inherent generation process. This enables effective ownership verification without the need for intrusive watermarks. The framework assesses the source model's conditions and evaluates a suspect model in two scenarios: white-box pipeline access and black-box access. The paper responds to rising concerns regarding intellectual property protection for hosted services and downloadable checkpoints, especially in instances of model leakage, unauthorized copying, or fine-tuning. Authored by an unnamed team, this research emphasizes the framework's non-invasive advantage, as it does not modify the model or its outputs. The study can be accessed at arxiv.org/abs/2608.11732.