知识卡片
扩散生成需定义完整采样器
去噪网络之外,还需指定终态初始分布、时间调度、各步均值或漂移、随机项及最终输出处理。有限步误差和网络误差都会影响生成分布。Heeger–Bergen同样从噪声迭代产生纹理,只构成结构类比,不替代这些概率定义。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 生成模型
原书出处 · 1 处
- · 查看原书依据
We call fθ a denoiser; it learns to remove a little bit of noise from an image. If we apply this denoiser over and over, starting with pure noise, the process should coalesce on a noise-less image that looks like one of our training examples. The steps to train a diffusion model are therefore as follows:
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:摘录中关于扩散生成需定义完整采样器的定义、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:We call fθ a denoiser; it learns to remove a little bit of noise from an image. If we apply this denoiser over and over, starting with pure noise, the process should coalesce on a noise-less image that looks like one of our training examples. The steps to train a diffusion model are therefore as follows: