Search papers, labs, and topics across Lattice.
This paper introduces iterative self-improving codebooks for safe autoregressive image generation, addressing the safety concerns inherent in generating images from discretized visual tokens. By leveraging the unified multimodal model's capabilities to identify unsafe images without human annotation, the authors construct harmful and safe image-text pairs to guide updates to the codebook. The method iteratively refines the codebook to eliminate harmful mappings while maintaining the quality of generated images, resulting in a safety-enhanced model that improves without external feedback.
Autoregressive image generation can be made significantly safer by iteratively refining codebooks based on the model's own judgment of harmful outputs.
Unlike diffusion-based models that operate in continuous latent spaces, autoregressive unified multimodal models produce images by sequentially predicting discretized visual tokens. These tokens are derived from a codebook that maps embeddings to quantized visual patterns. The language-like architecture enables unified multimodal models to effectively capture text conditional information for generation, making them promising for text-to-image tasks. This also raises an interesting question: how safe are the images generated in such an autoregressive way? In this work, we propose iterative self-improving codebooks for safe autoregressive generation. We leverage the understanding and judgment capabilities of the unified multimodal model itself to identify unsafe generated images without human annotation. Subsequently, the inherent representations in the codebook are fixed to eliminate harmful mappings. Our method comprises two steps: first, we use the unified model to identify unsafe generations and construct corresponding harmful and safe image-text pairs. These pairs are used to construct the Harmful Space and guide updates to the codebook, thereby eliminating harmful outputs. Second, we perform adaptive fine-tuning on the codebook within the harmless space using safe image-text pairs to ensure the quality of generated images. These two steps are repeated until no further improvement is observed, producing a safety-enhanced model codebook. Without additional external feedback, the safety of models is improved iteratively.
