Reproduce “Safety Without Semantic Disruptions: Editing-free Safe Image Generation via Context-preserving Dual Latent Reconstruction” on OpenResearch

About this paper

Jordan Vice, Naveed Akhtar, Mubarak Shah, Richard Hartley, Ajmal Mian

Researchers at the University of Western Australia, University of Melbourne, University of Central Florida, Australian National University, and Google developed an editing-free method for safe image generation in text-to-image models. This approach leverages context-preserving dual latent reconstruction to prevent semantic disruptions, achieving state-of-the-art safety on datasets like I2P while maintaining global visual context and generative diversity.

Key insights

• Model editing techniques cause measurable 'semantic disruptions' by shifting both unsafe and proximal benign concepts towards an unguided embedding space, compromising manifold integrity.
• It is possible to achieve high levels of content safety in generative models without directly editing the model's weights, thus avoiding collateral damage to semantic relationships.
• Maintaining global visual context during the early stages of the diffusion process is crucial for preventing unsafe content while preserving the overall scene composition and image coherence.

Summary by alphaXiv.