| （英） |
With the introduction of Low-Rank Adaptation (LoRA) in Stable Diffusion, fine-tuning on existing images has become more accessible, enabling the generation of diverse outputs. Alongside this progress, the protection of intellectual property rights has gained increasing importance, including copyright for image creators such as illustrators and photographers, portrait rights of individuals depicted, and trademark rights of characters. Against this backdrop, the research field of GenAI watermarking has recently attracted attention as a method that enables model training while preserving the quality of original images, yet intentionally degrades the quality of images generated by the resulting models.
Various approaches leveraging the characteristics of Stable Diffusion have been proposed for GenAI watermarking. In this study, we focus on the VAE encoder, which is responsible for mapping images into the latent space in Stable Diffusion. We propose a method that adjusts perturbations applied to the original image so that it is transformed into a perturbed image embedding a watermark within the latent space. The proposed method minimizes visual differences from the original image while maximizing latent-space discrepancies, thereby ensuring that perturbations emerge in generated images during LoRA training. |