← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6577 · 生成模型:GPT-6

语义编辑需要保留未变机制

理解等级:中级

因果语义编辑尝试改变一个生成因素,同时按模型保留其他背景并传播下游变化。这与只擦掉一块像素或任意移动潜在向量不同。若生成因素之间有因果关系,其后代也可能应变化,不能一律要求所有未指定因素完全固定。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      semantic editing, where rather than manipulating the pixels in the image, I manipulate some latent representation of the image corresponding to “glasses.” In effect, I pose the counterfactual question, “what would this image look like if the subject were not wearing glasses?” Figure 9.16 contrasts the original and “counterfactual” versions of the image.

    参考来源

    - 位置:《Causal AI》第09章《通用反事实推断算法》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Causal AI/_epub-src/OEBPS/Text/chapter-9.xhtml) - 结论依据:摘录中关于语义编辑需要保留未变机制的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:semantic editing, where rather than manipulating the pixels in the image, I manipulate some latent representation of the image corresponding to “glasses.” In effect, I pose the counterfactual question, “what would this image look like if the subject were not wearing glasses?” Figure 9.16 contrasts the original and “counterfactual” versions of the image.