A Position-Based Taxonomy of In-Generation Watermarking for Latent Diffusion Models

Generative artificial intelligence, especially Text-to-Image technology represented by latent diffusion models, has reshaped the landscape of digital content creation but has also raised deep concerns about false information, copyright ownership, and content traceability. Digital watermarking technology, as an imperceptible identification and traceability mechanism, has become a key technical means to achieve AI-generated content (AIGC) governance and copyright protection. This paper presents a systematic review of in-generation watermarking for latent diffusion models and develops a position-based taxonomy according to where watermarks are embedded in the generation pipeline and model architecture. Furthermore, we analyze the adversarial security threats faced by these watermarks and summarize corresponding defense strategies. Finally, we systematize key evaluation dimensions and identify the open challenges currently faced by the technology. This paper serves as a reference to facilitate future research in this field.