Figureure 2 · An overview of our method for constructing an iterative self-improvingFigure 2. An overview of our method for constructing an iterative self-improving safe codebook. (a) Leveraging the image generation and understanding capabilities of a unified multimodal model to construct paired harmful and safe image-text data. (b) Building harmful space from corresponding harmful and safe prompts and removing the codebook’s projection within this space. (c) Adaptively fine-tuning the codebook using safe image-text pairs.这张图概括 Safe Autoregressive Image Generation with 的整体方法流程。阅读时先看模块之间传递的训练信号,再看作者如何把目标拆成可优化的子问题。Figureure 3 · Impact of our method on the model’s inherent image generation and imagFigure 3. Impact of our method on the model’s inherent image generation and image understanding capabilities. Applying the safety codebook we constructed does not significantly degrade the model’s original performance.这张图/表用于判断 Safe Autoregressive Image Generation with 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致,而不是只看单个最高分。