<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[Funnyman_Ming的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/Funnyman_Ming</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; Funnyman_Ming]]></copyright><item><title><![CDATA[文生图相关论文汇总]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142437444</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142437444</guid><author>Funnyman_Ming</author><pubDate>Sun, 22 Sep 2024 20:20:57 +0800</pubDate><description><![CDATA[DDPM是"Denoising Diffusion Probabilistic Models"的缩写， 去噪扩散概率模型。]]></description><category></category></item><item><title><![CDATA[CTC loss 博客转载]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142428033</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142428033</guid><author>Funnyman_Ming</author><pubDate>Sun, 22 Sep 2024 02:31:10 +0800</pubDate><description><![CDATA[这时候我们就可以填表了，表中每个圆圈都是概率值，第一行第一列的数值就是，字符1是类别空格的概率，第二行第一列的数值就是，字符1是类别a的概率。模型预测结果是8*37的大小，换句话说就是，预测出来了8个字符，每个字符的37个类别的概率。一个合法路径就是一种预测结果为正确的情况，每一个合法路径上的概率相乘就是这个路径的概率。英文OCR，37个类别（26个小写字母+10个汉字+空格），最大输出长度8个字符。优化目标是最大化“所有合法路径概率的和”；]]></description><category></category></item><item><title><![CDATA[SVTR文字识别]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142426230</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142426230</guid><author>Funnyman_Ming</author><pubDate>Sun, 22 Sep 2024 02:13:09 +0800</pubDate><description><![CDATA[论文2.5中说的N nodes，就是输出的类别数量，英文37，中文6625，英文37说的是最简单的英文文字识别任务，不区分大小写，就是26个字母+10个数字+空格；高度下降为一半，宽度不降，是因为高度为度上采用stride=2，宽度上stride=1；就是一个窗口内的Global Mixing，局部的自注意，添加局部范围的信息；就是全局的自注意，类似与transformer，添加全局的信息；BN就是BatchNorm，LN就是layer norm；]]></description><category></category></item><item><title><![CDATA[transformer目标检测 DETR]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142385266</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142385266</guid><author>Funnyman_Ming</author><pubDate>Fri, 20 Sep 2024 23:42:35 +0800</pubDate><description><![CDATA[亮点：位置编码，encoder中加了，decoder中也加了；没用NMS，用的匈牙利匹配；两个注意力自注意力+交叉注意，交叉注意在decoder，object queries做的是qurey；object queries是固定100个；（在某些位置上查询有无物体以及类别）]]></description><category></category></item><item><title><![CDATA[Batch Normalization]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142407596</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/142407596</guid><author>Funnyman_Ming</author><pubDate>Fri, 20 Sep 2024 23:28:52 +0800</pubDate><description><![CDATA[（训练阶段滑动平均；推理阶段使用累计的整体均值和方差）两个可学习的参数，缩放和平移；降低初始权重的敏感性；加速收敛，又快又稳；避免梯度消失与爆炸；]]></description><category></category></item><item><title><![CDATA[让TA说话 - 图片、动画、语音相结合]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136453177</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136453177</guid><author>Funnyman_Ming</author><pubDate>Thu, 07 Mar 2024 00:13:22 +0800</pubDate><description><![CDATA[如何让图片开口说话，或者让视频里的人换一段话？开源的技术已经做的不错了。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion 光影艺术字]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136374160</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136374160</guid><author>Funnyman_Ming</author><pubDate>Thu, 29 Feb 2024 17:40:04 +0800</pubDate><description><![CDATA[使用sd制作光影文字效果]]></description><category></category></item><item><title><![CDATA[Stable Diffusion 相关网站总结]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136241925</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136241925</guid><author>Funnyman_Ming</author><pubDate>Thu, 22 Feb 2024 19:45:50 +0800</pubDate><description><![CDATA[civitAI，sd开源社区，大部分开源模型都从这里下载：lightflow，一键复制流，有很多大神灵活使用sd的例子：]]></description><category></category></item><item><title><![CDATA[Stable Diffusion 优质图生成技巧]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136233282</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136233282</guid><author>Funnyman_Ming</author><pubDate>Thu, 22 Feb 2024 19:26:54 +0800</pubDate><description><![CDATA[在之前的博客中，为了追求功能的准确，我大都没开各种修复功能，并且都是使用很低的分辨率，这就导致以前博客中的图都不美观。今天专门分享一次如何出优质图、高清图、精致图、看上去就很美的图。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion 插件AnimateDiff]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136194394</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/136194394</guid><author>Funnyman_Ming</author><pubDate>Wed, 21 Feb 2024 11:16:52 +0800</pubDate><description><![CDATA[AnimateDiff可以制作动图，以插件的方式集成到sd的webUI，本文介绍了AnimateDiff的安装与使用。]]></description><category></category></item><item><title><![CDATA[float浮点数转音频文件wav]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135817607</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135817607</guid><author>Funnyman_Ming</author><pubDate>Wed, 24 Jan 2024 11:27:14 +0800</pubDate><description><![CDATA[java要是完成类似的功能就比较复杂。python将numpy.ndarray转为音频文件wav，只需要几行代码。]]></description><category></category></item><item><title><![CDATA[paddle系列报错 Segmentation fault]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135773655</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135773655</guid><author>Funnyman_Ming</author><pubDate>Tue, 23 Jan 2024 16:02:05 +0800</pubDate><description><![CDATA[这行代码的意思就是配置LD_LIBRARY_PATH这个环境变量，这是临时配置，关闭控制台后就失效了，需要长久生效的配置需要自己改配置表。paddlepaddle_gpu自从2.1.3之后，在一些linux环境中部署总会报错。paddle的cpu版本并不会报错。但是，经过我的实验，python代码中配置，依然会报错。我主要是参考了这两个博客，我简化了指令。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion ControlNet 与艺术字（2）]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135508832</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135508832</guid><author>Funnyman_Ming</author><pubDate>Wed, 10 Jan 2024 19:25:28 +0800</pubDate><description><![CDATA[如果使用的字体直线多、棱角分明，生成的图中字上的内容和提示词差距越大。比如Office带的娃娃体，可以看到字上面一点儿花的元素都没有，如下图。在笔直的字体上做了一些其他的尝试，也很难把提示内容加进去，我尝试不同的权重组合和提示词，发现最多只带有很少一部分的提示词内容。具体的配置参数写在了最后面，配置主要参考了博主。反向词：NSFW,worst quality,low quality,hand,people,提示词：Beautiful flowers in the sky,8k,realistic,]]></description><category></category></item><item><title><![CDATA[Stable Diffusion UNet]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135503726</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135503726</guid><author>Funnyman_Ming</author><pubDate>Wed, 10 Jan 2024 14:59:14 +0800</pubDate><description><![CDATA[UNet是sd的核心模块，值得注意的是UNet预测的是噪声，整个生成过程就是一个去噪的过程。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion CLIP]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135457953</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135457953</guid><author>Funnyman_Ming</author><pubDate>Mon, 08 Jan 2024 16:48:33 +0800</pubDate><description><![CDATA[sd主要包含了3大块，VAE、CLIP和UNet。其中CLIP（Contrastive Language-Image Pre-training）主要是用来将提示词转为tokens，通俗的说，就是把文字转化为可以模型可以理解的信息。这些语义信息是77*768大小的向量，在UNet阶段一步一步嵌入进去，这样生成的图片就能附带提示词的内容，嵌入方式为cross attention。对输入text提取text embeddings的方式可以是其他的，目前sd主要使用OpenAI的CLIP。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion VAE]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135414034</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135414034</guid><author>Funnyman_Ming</author><pubDate>Fri, 05 Jan 2024 18:20:10 +0800</pubDate><description><![CDATA[Stable Diffusion 中VAE的编码和解码过程的直观展示。]]></description><category></category></item><item><title><![CDATA[Stable Diffusion ControlNet 与艺术字]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135390468</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/135390468</guid><author>Funnyman_Ming</author><pubDate>Thu, 04 Jan 2024 17:49:56 +0800</pubDate><description><![CDATA[sd画的图可控性并不高，尤其是文字，比如我们想让sd生成一个带有汉字“王炸”的条幅，sd会生成条幅，但是汉字都是有形体，但是内容是不对的，如图：ControlNet为解决这一问题作出了突出贡献。sd是一种UNet，包含编码和解码的过程，文生图的过程就是字转为带有语义信息的token，并逐步附加到sd的decoder过程中。ControlNet同样也是将输入转为向量，逐步添加到生成过程。]]></description><category></category></item><item><title><![CDATA[Anaconda 3 下的 Spyder 的汉化报错处理]]></title><link>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/82590916</link><guid>https://blog.youkuaiyun.com/Funnyman_Ming/article/details/82590916</guid><author>Funnyman_Ming</author><pubDate>Mon, 10 Sep 2018 16:19:35 +0800</pubDate><description><![CDATA[Anaconda 3 下的 Spyder 的汉化

参照原文博客：http://www.lizenghai.com/archives/523.html

汉化包下载：https://github.com/kingmo888/Spyder_Simplified_Chinese

本人在汉化过程中遇见几个问题，特与大家分享经验。

1.运行错误

在参照原文博主的方法进行汉化中遇见问题，如下：



...]]></description><category></category></item></channel></rss>