In-Generation Watermarking for Latent Diffusion Models: A Perceptual Guidance Mechanism via Texture and JND

AI-generated images are currently facing severe challenges regarding copyright ownership and unauthorized usage. In this paper, we propose T-JIG, an endogenous watermarking method targeting latent diffusion models (LDM). This method incorporates a JND (Just Noticeable Distortion) perceptual correction module along with a texture loss module, which encourages the distribution of watermark information in complex texture regions, thereby reducing the risk of artifact formation in visually sensitive areas. Additionally, to ensure the method's applicability to video watermark embedding tasks, an extra video compression attack is introduced within the noise layer.