← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8987.d · 生成模型:GPT-6

预测噪声需配套缩放公式

理解等级:中级

若针对单步xt=√(1−β)xt−1+√β ε估计噪声ε̂,代数恢复应为(xt−√β ε̂)/√(1−β),并限定β;实际模型还涉及条件估计与反向采样。书中32.48写成xt+ε且未缩放,不能直接作为通用扩散更新式。

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

      One useful trick for training diffusion models is to reparameterize the learning problem as predicting the noise rather than the signal [208]: (32.47) (32.48) One way to look at diffusion models is that we want to learn a mapping from pure noise (e.g., z ∼ N (0, I)) to data. It is very hard to figure out how to create structured data out of noise,

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《生成模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c032.xhtml) - 结论依据:摘录中关于预测噪声需配套缩放公式的定义、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:One useful trick for training diffusion models is to reparameterize the learning problem as predicting the noise rather than the signal [208]: (32.47) (32.48) One way to look at diffusion models is that we want to learn a mapping from pure noise (e.g., z ∼ N (0, I)) to data. It is very hard to figure out how to create structured data out of noise,