[{"content":"论文：Ho, Jain, and Abbeel. Denoising Diffusion Probabilistic Models, 2020.\n一句话概括：用固定的前向过程逐步破坏数据，再训练神经网络学习逆转这一过程。\n这篇论文解决什么问题 # DDPM 研究的是如何学习复杂的图像分布。它不直接让网络一次生成完整图像，而是把生成过程拆成许多简单的去噪步骤：从高斯噪声开始，每一步去掉少量噪声，最终得到符合数据分布的样本。\n这种设计的关键价值在于，困难的“从无到有”被改写成了一系列更稳定的局部去噪问题。\n前向过程：如何把图像变成噪声 # 给定干净图像 \\(x_0\\)，前向过程在每一步加入少量高斯噪声：\n\\[ q(x_t\\mid x_{t-1}) = \\mathcal{N}(x_t; \\sqrt{1-\\beta_t}x_{t-1}, \\beta_t I). \\]记 \\(\\alpha_t=1-\\beta_t\\)，\\(\\bar\\alpha_t=\\prod_{s=1}^{t}\\alpha_s\\)，可以直接从 \\(x_0\\) 采样任意时刻的 \\(x_t\\)：\n\\[ x_t=\\sqrt{\\bar\\alpha_t}x_0+\\sqrt{1-\\bar\\alpha_t}\\epsilon,\\quad \\epsilon\\sim\\mathcal N(0,I). \\]这条公式非常重要：训练时无需真的一步步加噪，只需随机采样时间步 \\(t\\) 和噪声 \\(\\epsilon\\)，便能构造训练样本。\n反向过程：网络究竟预测什么 # 真实的反向分布未知，因此使用参数化高斯分布 \\(p_\\theta(x_{t-1}\\mid x_t)\\) 逼近。论文采用 U-Net 预测加入到 \\(x_t\\) 中的噪声 \\(\\epsilon\\)。常见的简化目标是：\n\\[ L_{simple}=\\mathbb E_{t,x_0,\\epsilon}\\left[\\lVert\\epsilon-\\epsilon_\\theta(x_t,t)\\rVert^2\\right]. \\]直觉上，网络看到带噪图像和时间步，需要判断“其中哪一部分是噪声”。采样时反复调用这个网络，就能逐渐恢复图像结构。\n阅读后的三个关键认识 # 时间步是噪声强度的坐标。 同一个去噪网络需要处理从轻微扰动到接近纯噪声的多种状态，因此时间嵌入不可缺少。 训练与采样的代价不对称。 训练可以一次构造任意 \\(x_t\\)，而原始采样仍需执行完整的马尔可夫链，这也是后续加速工作的主要切入点。 扩散模型与图像复原天然相容。 去噪、去模糊、超分辨率等任务都可被理解为在观测约束下恢复干净图像分布。 和图像复原的连接 # 无条件 DDPM 学习的是图像先验；图像复原则需要把退化观测 \\(y\\) 加入反向过程。后续方法通常通过条件网络、引导项或数据一致性约束，让每一步去噪既符合自然图像分布，也不偏离输入观测。\n仍然值得追问 # 噪声预测、数据预测与 velocity prediction 在复原任务中有什么实际差异？ 如何在减少采样步数的同时维持高频细节？ 面对未知退化时，生成先验会带来恢复能力还是不可信的幻觉？ 原论文：Denoising Diffusion Probabilistic Models\n","date":"2026-07-23","externalUrl":null,"permalink":"/papers/ddpm/","section":"论文阅读","summary":"DDPM 为什么可以从高斯噪声中生成图像？这篇笔记从前向加噪、反向去噪和噪声预测目标三个部分建立直觉。","title":"读懂 DDPM：从加噪过程到噪声预测","type":"papers"},{"content":"论文：Zamir et al. Restormer: Efficient Transformer for High-Resolution Image Restoration, CVPR 2022.\n一句话概括：把注意力的主要计算从空间维度转移到通道维度，使 Transformer 能够处理高分辨率复原任务。\n为什么普通自注意力不适合图像复原 # 图像复原通常保持输入分辨率，特征图中的 token 数量 \\(HW\\) 很大。标准自注意力需要构造 \\(HW\\times HW\\) 的相关矩阵，复杂度随像素数量平方增长。对于高分辨率图像，这一成本很快变得不可接受。\n另一方面，卷积虽然高效，却受限于固定的局部感受野。Restormer 的目标是在二者之间找到平衡：保留全局建模能力，同时让计算量随空间尺寸近似线性增长。\n整体结构 # Restormer 采用四层 U-Net 式编码器—解码器结构。编码阶段逐步降低空间分辨率、增加通道数；解码阶段恢复分辨率，并使用跳跃连接传递高分辨率细节。最高分辨率阶段末尾还有 refinement blocks，用于进一步修正输出。\n真正决定模型特性的，是每个 Transformer block 中的两个模块：MDTA 和 GDFN。\nMDTA：在通道维度计算注意力 # Multi-Dconv Head Transposed Attention 先用 \\(1\\times1\\) 卷积聚合通道信息，再用 \\(3\\times3\\) depth-wise 卷积引入局部上下文。与标准注意力不同，它计算的是通道之间的相关性，而不是所有空间位置之间的相关性。\n这样得到的注意力矩阵规模与通道数有关，避开了 \\((HW)^2\\) 的空间复杂度。可以把它理解为：模型先在整幅图像上汇总每个通道的响应，再学习哪些特征通道应当相互增强或抑制。\nGDFN：带门控的前馈网络 # Gated-Dconv Feed-Forward Network 将特征投影成两条分支，其中一条经过 GELU 激活，两条分支逐元素相乘：\n\\[ \\hat X = W_p^0\\,\\mathrm{GELU}(W_d^1W_p^1X)\\odot(W_d^2W_p^2X)+X. \\]门控机制让网络选择性地保留有用特征；depth-wise 卷积则补充了前馈网络原本缺少的局部空间建模。这对边缘、纹理和微小退化尤其重要。\n为什么它对复原任务有效 # 全局关系：MDTA 能利用远距离区域帮助判断局部退化。 局部归纳偏置：两个核心模块都包含 depth-wise convolution，保留了卷积对局部结构的优势。 多尺度结构：编码器—解码器同时处理语义上下文与高分辨率细节。 任务通用性：论文在去雨、运动与散焦去模糊、合成及真实图像去噪等任务上进行验证。 阅读后的判断 # Restormer 的贡献不只是“把 Transformer 用于复原”，而是重新选择了注意力发生的维度。它说明全局建模并不意味着必须显式比较每一对像素；通道统计同样可以传递全局信息。\n它也给 Diffusion-based restoration 一个直接启发：如果去噪网络需要反复运行，高分辨率特征上的注意力成本会被采样步数放大，因此高效的全局模块尤其重要。\n仍然值得追问 # 通道注意力在多大程度上能替代显式的空间长程关系？ 同一套结构跨多种退化有效，但面对真实混合退化时是否仍需专门建模？ 将 Restormer block 用作扩散模型去噪器时，质量与采样速度之间如何权衡？ 原论文：Restormer: Efficient Transformer for High-Resolution Image Restoration\n","date":"2026-07-20","externalUrl":null,"permalink":"/papers/restormer/","section":"论文阅读","summary":"Restormer 不在空间位置之间构建平方复杂度的注意力，而是在通道维度建模全局关系，并用门控前馈网络保留局部结构。","title":"Restormer：高分辨率图像复原中的高效 Transformer","type":"papers"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/authors/ajay-jain/","section":"Authors","summary":"","title":"Ajay Jain","type":"authors"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/authors/","section":"Authors","summary":"","title":"Authors","type":"authors"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/","section":"BOLIN的个人博客","summary":"","title":"BOLIN的个人博客","type":"page"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/categories/","section":"Categories","summary":"","title":"Categories","type":"categories"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/tags/ddpm/","section":"Tags","summary":"","title":"DDPM","type":"tags"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/categories/diffusion/","section":"Categories","summary":"","title":"Diffusion","type":"categories"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/tags/diffusion/","section":"Tags","summary":"","title":"Diffusion","type":"tags"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/authors/jonathan-ho/","section":"Authors","summary":"","title":"Jonathan Ho","type":"authors"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/authors/pieter-abbeel/","section":"Authors","summary":"","title":"Pieter Abbeel","type":"authors"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/tags/","section":"Tags","summary":"","title":"Tags","type":"tags"},{"content":"这里收录论文的结构化阅读笔记。每篇文章尽量回答三个问题：论文解决了什么、方法为什么有效，以及它对图像复原研究有什么启发。\n","date":"2026-07-23","externalUrl":null,"permalink":"/papers/","section":"论文阅读","summary":"这里收录论文的结构化阅读笔记。每篇文章尽量回答三个问题：论文解决了什么、方法为什么有效，以及它对图像复原研究有什么启发。\n","title":"论文阅读","type":"papers"},{"content":"","date":"2026-07-23","externalUrl":null,"permalink":"/tags/%E8%AE%BA%E6%96%87%E9%98%85%E8%AF%BB/","section":"Tags","summary":"","title":"论文阅读","type":"tags"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/aditya-arora/","section":"Authors","summary":"","title":"Aditya Arora","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/fahad-shahbaz-khan/","section":"Authors","summary":"","title":"Fahad Shahbaz Khan","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/categories/image-restoration/","section":"Categories","summary":"","title":"Image Restoration","type":"categories"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/tags/image-restoration/","section":"Tags","summary":"","title":"Image Restoration","type":"tags"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/ming-hsuan-yang/","section":"Authors","summary":"","title":"Ming-Hsuan Yang","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/munawar-hayat/","section":"Authors","summary":"","title":"Munawar Hayat","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/salman-khan/","section":"Authors","summary":"","title":"Salman Khan","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/authors/syed-waqas-zamir/","section":"Authors","summary":"","title":"Syed Waqas Zamir","type":"authors"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/tags/transformer/","section":"Tags","summary":"","title":"Transformer","type":"tags"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/tags/%E5%8E%BB%E6%A8%A1%E7%B3%8A/","section":"Tags","summary":"","title":"去模糊","type":"tags"},{"content":"","date":"2026-07-20","externalUrl":null,"permalink":"/tags/%E5%8E%BB%E5%99%AA/","section":"Tags","summary":"","title":"去噪","type":"tags"},{"content":"这里将整理 All-in-One Image Restoration 方向的论文，包括退化表征、任务提示、混合专家、统一预训练和未知退化泛化等主题。\n新的阅读笔记会持续补充到这里。\n","externalUrl":null,"permalink":"/topics/all-in-one-restoration/","section":"Topics","summary":"这里将整理 All-in-One Image Restoration 方向的论文，包括退化表征、任务提示、混合专家、统一预训练和未知退化泛化等主题。\n","title":"All-in-One Restoration","type":"topics"},{"content":"","externalUrl":null,"permalink":"/series/","section":"Series","summary":"","title":"Series","type":"series"},{"content":"","externalUrl":null,"permalink":"/topics/","section":"Topics","summary":"","title":"Topics","type":"topics"},{"content":"这里将整理 Super-Resolution 方向的论文，关注退化建模、感知质量、真实场景超分辨率，以及 Diffusion 在超分辨率中的应用。\n新的阅读笔记会持续补充到这里。\n","externalUrl":null,"permalink":"/topics/super-resolution/","section":"Topics","summary":"这里将整理 Super-Resolution 方向的论文，关注退化建模、感知质量、真实场景超分辨率，以及 Diffusion 在超分辨率中的应用。\n","title":"超分辨率","type":"topics"},{"content":"这里按发布时间整理所有论文笔记与技术分享。\n","externalUrl":null,"permalink":"/archives/","section":"归档","summary":"这里按发布时间整理所有论文笔记与技术分享。\n","title":"归档","type":"archives"}]