Safe-VAR introduces the first watermarking technique specifically designed for Visual Autoregressive (VAR) models used in text-to-image generation, addressing a critical security gap in AI safety.

• Adapts watermarking to the sequential nature of VAR models, unlike previous methods built for diffusion models
• Provides robust protection against common image manipulations and perturbation attacks
• Enables origin verification of AI-generated content to prevent misuse and unauthorized redistribution
• Maintains high image quality while embedding invisible watermarks

This research is crucial for the security community as generative AI becomes more mainstream, offering a practical solution to identify AI-generated content, protect intellectual property, and establish accountability in AI creation.

Safe-VAR: Safe Visual Autoregressive Model for Text-to-Image Generative Watermarking