← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8987.c · 生成模型:GPT-6

均方去噪不能唯一逆转噪声

理解等级:高级

噪声损坏通常多对一,给定xt不能精确知道原先是哪一个xt−1。平方损失预测的是相应条件均值,而非每条路径的确定逆。要从噪声恢复完整数据分布,需定义合适的反向转移或其他采样动力学,不能只保证反复应用均值预测就正确抽样。

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

      In algorithm 32.2 we train the denoiser using a loss L, which could be the L2 distance. In many diffusion models, fθ is instead trained to output the parameters (mean and variance) as a Gaussian density model fit to the data. This formulation yields useful probabilistic interpretations (in fact, such a diffusion model can be framed as a variational autoencoder, which we will see in

    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:摘录中关于均方去噪不能唯一逆转噪声的定义、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:In algorithm 32.2 we train the denoiser using a loss L, which could be the L2 distance. In many diffusion models, fθ is instead trained to output the parameters (mean and variance) as a Gaussian density model fit to the data. This formulation yields useful probabilistic interpretations (in fact, such a diffusion model can be framed as a variational autoencoder, which we will see in