<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>AIGC - Naifan Li's Blog</title><link>https://blog.omagiclee.com/categories/aigc/</link><description>AIGC - Naifan Li's Blog</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><lastBuildDate>Thu, 01 Oct 2026 10:00:00 +0800</lastBuildDate><atom:link href="https://blog.omagiclee.com/categories/aigc/index.xml" rel="self" type="application/rss+xml"/><item><title>ComfyUI 设计架构：从节点画布到模型执行引擎</title><link>https://blog.omagiclee.com/posts/aigc/comfyui/architecture/</link><pubDate>Thu, 01 Oct 2026 10:00:00 +0800</pubDate><guid>https://blog.omagiclee.com/posts/aigc/comfyui/architecture/</guid><description><![CDATA[<p>在 ComfyUI 的画布上，把模型加载、文本编码、采样和图片保存几个节点连起来，就可以生成一张图片。但这条连线背后还有一套完整的软件系统：浏览器把工作流转换成执行图，Python 服务负责校验和排队，执行器按数据依赖调用节点，模型运行层再完成张量计算与设备内存管理。</p>
<p><strong>理解 ComfyUI，可以沿着“可视化编排 → 图执行 → 模型计算”这三个层次展开。</strong> 本文先画出系统全貌，再沿着一次生成请求，解释每个模块的职责。</p>
<p>本文以开源、自托管 ComfyUI 的核心执行链路为范围，资料核对日期为 <strong>2026-10-01</strong>。下列图是依据官方文档与源码整理的逻辑视图，省略了用户管理、资产管理等辅助模块；不同安装版本的具体实现可能有所差异。</p>]]></description></item><item><title>Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition</title><link>https://blog.omagiclee.com/posts/aigc/qwen/qwen-image-layered/</link><pubDate>Tue, 23 Dec 2025 14:35:16 +0800</pubDate><guid>https://blog.omagiclee.com/posts/aigc/qwen/qwen-image-layered/</guid><description><![CDATA[<p><i class="fas fa-building fa-fw" aria-hidden="true"></i><span style="color:gray">QwenLM</span>
<i class="fas fa-file-pdf fa-fw" aria-hidden="true"></i><a href="https://arxiv.org/abs/2511.17374" target="_blank" rel="noopener noreffer ">arXiv 2511.17374</a>
<i class="fab fa-github fa-fw" aria-hidden="true"></i><a href="https://github.com/QwenLM/Qwen-Image-Layered" target="_blank" rel="noopener noreffer ">QwenLM/Qwen-Image-Layered</a>
</p>
<h2 id="motivation">Motivation</h2>
<h2 id="contribution">Contribution</h2>
<h2 id="method">Method</h2>
<h3 id="model-architecture">Model Architecture</h3>
<h2 id="experiment">Experiment</h2>
<h2 id="reference">Reference</h2>
<h2 id="question">Question</h2>]]></description></item></channel></rss>