<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[yinglang19941010的专栏]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/yinglang19941010</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; yinglang19941010]]></copyright><item><title><![CDATA[开发基础（一）UV]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/162276571</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/162276571</guid><author>yinglang19941010</author><pubDate>Wed, 24 Jun 2026 17:20:03 +0800</pubDate><description><![CDATA[uv 会自动创建并管理 Python 环境，这是它相比传统 pip + venv 工作流的核心优势之一。]]></description><category></category></item><item><title><![CDATA[【】提示词工程的迭代SOP]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/157763868</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/157763868</guid><author>yinglang19941010</author><pubDate>Mon, 09 Feb 2026 10:14:34 +0800</pubDate><description><![CDATA[摘要 本文提出了基于Agent的任务处理优化策略与标准流程。在策略层面，包括：1）输入预处理（图片裁剪/视觉分析）；2）提示词优化（关键词替换、约束位置调整、任务拆解）；3）Multi-Agent架构（并行处理子任务/串行提高准确率）；4）分支Agent（针对不同数据模式）。在执行流程上，建议分两步验证：先用3个案例测试模型可行性，再通过30+案例集验证可用率。这些方法通过调整输入、优化提示词和智能体协作，能有效提升复杂任务的完成质量。]]></description><category></category></item><item><title><![CDATA[【VLM模型】(一) 轻量化模型]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/156107910</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/156107910</guid><author>yinglang19941010</author><pubDate>Mon, 22 Dec 2025 11:27:40 +0800</pubDate><description><![CDATA[截至2025-12-19，MiniCPM-V系列（以3B级为主力，最新到4.5）在轻量级VLM（≤7B）中属于中文友好、OCR/文档强、推理效率高的第一梯队，综合性价比与落地友好度突出；与Qwen2.5-VL、SmolVLM等相比，各有侧重，适合不同场景。MiniCPM-V支持 int4, GGUF。]]></description><category></category></item><item><title><![CDATA[【生成模型】【强化学习】（一）RLHF & DPO]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/154008800</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/154008800</guid><author>yinglang19941010</author><pubDate>Sat, 15 Nov 2025 07:48:57 +0800</pubDate><description><![CDATA[大型语言模型（如 GPT 系列）在训练后，往往需要“对齐”（alignment），让它们生成更符合人类偏好的输出。比如，你问 AI 一个问题，它应该给出有帮助、礼貌的回答，而不是胡说八道或有害的内容。传统的对齐方法是 RLHF（Reinforcement Learning from Human Feedback，从人类反馈中强化学习）：Step1：收集人类反馈数据。通常是给模型一个提示（prompt，比如“解释量子力学”），模型生成多个回答，然后人类标注哪个更好或者给回答打分s。]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（五）Qwen-Image-Edit：diffusers中LoRA加载与分析]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/154349012</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/154349012</guid><author>yinglang19941010</author><pubDate>Thu, 06 Nov 2025 14:17:18 +0800</pubDate><description><![CDATA[peft中的LoRA层对象（BaseTunerLayer的子类，比如Linear替换成peft.tuners.lora.layer.Linear），并将原本的Linear对象放到新对象的base_layer这个成员变量下，而peft.tuners.lora.layer.Linear调用时会根据激活的adapter自动应用lora。），会发现代码中既没有创建LoRA层，也没有在推理的时候调用LoRA层，那到底是什么时候调用的LoRA参数的呢？在diffsuers阅读某个pipeline的代码（如。]]></description><category></category></item><item><title><![CDATA[【开发】Git处理分支的指令]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/154453525</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/154453525</guid><author>yinglang19941010</author><pubDate>Wed, 05 Nov 2025 21:29:30 +0800</pubDate><description><![CDATA[是 git fetch + git merge 的组合命令：先拉取远程分支的最新代码（fetch），再将其合并到当前本地分支（默认用 merge 方式）可以将 pull 的合并方式改为变基（即 git fetch + git rebase），既同步远程代码，又保持历史线性，适合个人分支使用。-i表示手动修改提交记录，例如对弹出来的内容进行如下修改：D2改为fixup。会在dev分支上创建一个新commit节点，master上无新节点。]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（四）Qwen-Image-Edit]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/153274640</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/153274640</guid><author>yinglang19941010</author><pubDate>Tue, 21 Oct 2025 11:53:22 +0800</pubDate><description><![CDATA[step蒸馏模型：https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Distill-Full。]]></description><category></category></item><item><title><![CDATA[【环境安装常见问题】]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/150098758</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/150098758</guid><author>yinglang19941010</author><pubDate>Sat, 09 Aug 2025 23:48:57 +0800</pubDate><description><![CDATA[安装make（机器不能获取最新版的make）需要torch>=2.5。]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（三）Qwen-Image]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149930264</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149930264</guid><author>yinglang19941010</author><pubDate>Wed, 06 Aug 2025 10:37:29 +0800</pubDate><description><![CDATA[step蒸馏模型：https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Distill-Full。Text encoder是QwenVL，理论上可以突破token长度的限制；如果没有DASHSCOPE_API_KEY的话，注释掉rewrite模块。VAE使用的是Wan VAE。可能使用即梦做了数据蒸馏。需要torch>=2.5。]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（三）视频生成Wan2.2+lightx]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149887532</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149887532</guid><author>yinglang19941010</author><pubDate>Mon, 04 Aug 2025 13:18:03 +0800</pubDate><description><![CDATA[Wan 2.2视频生成模型在ComfyUI中的实现]]></description><category></category></item><item><title><![CDATA[【生成模型】Flux学习（一）推理代码]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149499588</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149499588</guid><author>yinglang19941010</author><pubDate>Mon, 21 Jul 2025 10:56:41 +0800</pubDate><description><![CDATA[]]></description><category></category></item><item><title><![CDATA[【VS Code】VS Code SFTP 插件配置]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149423872</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149423872</guid><author>yinglang19941010</author><pubDate>Thu, 17 Jul 2025 16:04:22 +0800</pubDate><description><![CDATA[最近训练IntenVL多模态模型的时候发现，模型推理时候的显存/速度占用明显高于训练时。]]></description><category></category></item><item><title><![CDATA[【生成模型】生成模型速度调研]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149332752</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149332752</guid><author>yinglang19941010</author><pubDate>Tue, 15 Jul 2025 10:17:36 +0800</pubDate><description><![CDATA[Swin-Transformer是一种分窗处理策略，降低空间上attention的复杂度；后面没有被广泛应用，是因为Deformable DETR这种Attetnion降低了复杂度；]]></description><category></category></item><item><title><![CDATA[【生成模型】【ComfyUI】一些重要的节点]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149097938</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149097938</guid><author>yinglang19941010</author><pubDate>Thu, 03 Jul 2025 17:05:53 +0800</pubDate><description><![CDATA[comfyui中一些常用的重要节点]]></description><category></category></item><item><title><![CDATA[【生成模型】【ComfyUI】一些环境问题]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/149020055</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/149020055</guid><author>yinglang19941010</author><pubDate>Mon, 30 Jun 2025 12:08:05 +0800</pubDate><description><![CDATA[在使用comfyui-easy-use中的【Image Remove Bg】节点，报错如下。检查环境，一般是cupy的版本和自己的cuda（11.8）不匹配。]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（一）视频生成Wan2.1速度质量简单评测]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/148345336</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/148345336</guid><author>yinglang19941010</author><pubDate>Fri, 27 Jun 2025 10:55:32 +0800</pubDate><description><![CDATA[VACE模型除了支持基本的文本和图像作为控制条件输入，还支持将视频和视频mask作为输入控制。简单理解就是视频版本的ControlNet和Flux-Fill。与图片生成的结构很像，VAE encode->Diffusion->VAE decode的形式，只是多了一个时间维度(T+1, H, W, C)]]></description><category></category></item><item><title><![CDATA[【生成模型】【模型介绍】（二）图像编辑 主体驱动 光照调整]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/148444390</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/148444390</guid><author>yinglang19941010</author><pubDate>Fri, 06 Jun 2025 14:58:21 +0800</pubDate><description><![CDATA[难点：（GPT4的长处）]]></description><category></category></item><item><title><![CDATA[【生成模型】视频生成论文调研]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/148423106</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/148423106</guid><author>yinglang19941010</author><pubDate>Wed, 04 Jun 2025 11:37:24 +0800</pubDate><description><![CDATA[基础模型上层模型。]]></description><category></category></item><item><title><![CDATA[【生成模型】AI生成API站点介绍]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/148396867</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/148396867</guid><author>yinglang19941010</author><pubDate>Tue, 03 Jun 2025 14:12:57 +0800</pubDate><description><![CDATA[]]></description><category></category></item><item><title><![CDATA[【生成模型】【基础知识】CFG与CFG蒸馏]]></title><link>https://blog.youkuaiyun.com/yinglang19941010/article/details/148170935</link><guid>https://blog.youkuaiyun.com/yinglang19941010/article/details/148170935</guid><author>yinglang19941010</author><pubDate>Fri, 23 May 2025 16:47:06 +0800</pubDate><description><![CDATA[是生成模型推理时的一种操作，目的是使的生成图像与输入的text prompt更接近。]]></description><category></category></item></channel></rss>