Follow me on Civitai <a target="_blank" rel="ugc" href="https://civarchive.com/user/lonecatone23">https://civarchive.com/user/lonecatone23</a>Instagram: <a target="_blank" rel="ugc" href="http://synth.studio">synth.studio</a>.models <a target="_blank" rel="ugc" href="https://www.instagram.com/synth.studio.models/">https://www.instagram.com/synth.studio.models/</a>This is many of hours of work. If you enjoy it, please 👍 like, 💬 comment, and feel free to ⚡ tip 😉<h3 id="buy-me-a-https:ko-fi.comlonecatone">Buy me a ☕ <a target="_blank" rel="ugc" href="https://ko-fi.com/lonecatone">https://ko-fi.com/lonecatone</a></h3><hr />Node Repo: <a target="_blank" rel="ugc" href="https://github.com/lonecatone23/ComfyUI_LC_Vision_nodes">https://github.com/lonecatone23/ComfyUI_LC_Vision_nodes</a><hr /><h3 id="lc-vision-why-this-pack-doesn't-waste-your-time">LC Vision 🔬📝🎥 — Why This Pack Doesn't Waste Your Time</h3>Introduction: Getting a vision-capable llama.cpp build running is normally a multi-hour fight. Qwen-VL support <code>Qwen3VLChatHandler</code> / <code>Qwen25VLChatHandler</code>) never got merged into upstream <code>llama-cpp-python</code>. It only lives in JamePeng's fork, shipped as raw GitHub Release wheels, one per exact Python version + CUDA version + OS combination. Get the match wrong and you get a bare import error, not a helpful one. This pack does that matching for you, automatically, before you've even opened a workflow.⚠️ <a target="_blank" rel="ugc" href="http://install.py">install.py</a> builds itself- Runs on its own the moment ComfyUI-Manager installs or updates this pack. You run nothing by hand.- First checks whether a vision-capable <code>llama_cpp</code> is already importable. If it is, it does nothing and gets out of your way.- If not, it reads your exact Python tag, OS, and CUDA version straight off your own machine, pulls JamePeng/llama-cpp-python's real GitHub Releases, and installs the one wheel that matches (rounding down to the closest CUDA version when there's no exact match).- Snapshots <code>pip freeze</code> before touching anything, and only ever steps in on a real conflict <code>pip check</code> reports afterward, and only when <code>llama-cpp-python</code> itself is the package at fault, never someone else's pin. Everything else in your environment is left alone.- Safe to re-run on every update. Idempotent by design, not by luck.How the Loader actually loads a model- <code>model_name</code> lists every <code>.gguf</code> + matching <code>mmproj</code> pair already sitting in <code>models/LLM</code>, plus a curated set of one-click downloads prefixed <code>⬇ Download:</code>.- Pick a <code>⬇ Download:</code> entry and it pulls straight from HuggingFace on first use, then lands in that same folder. Next time it's just a normal option in the list.- 💡 What's curated right now- Qwen3-VL-4B-abliterated and Qwen3-VL-8B-abliterated, each in Q8_0 (smaller, faster) and f16 (full precision). Four total, hand-picked instead of dumping an entire upstream catalog on you.- 4B is the video/Moviemaker pick, more headroom for multi-frame batches. 8B is the image/Caption pick, more fidelity.- mmproj always downloads at f16 regardless of the main model's quant. Vision quality leans on mmproj precision more than the LLM weights do.- Once loaded, it stays loaded. ComfyUI's own node caching skips re-running the Loader entirely as long as its inputs haven't changed, same as a checkpoint loader. No reload tax between every Caption, Enhancer, or Moviemaker call.Sell it: other vision-node packs hand you a requirements.txt and a wiki page telling you to go find the right wheel yourself. This one ships the fork detection, the CUDA matching, and the download baked in. Drop the node, pick a model, and it runs.<hr />关注我的 Civitai 主页 点击这里 <a target="_blank" rel="ugc" href="https://civarchive.com/user/lonecatone23">https://civarchive.com/user/lonecatone23</a>Instagram：<a target="_blank" rel="ugc" href="http://synth.studio">synth.studio</a>.models <a target="_blank" rel="ugc" href="https://www.instagram.com/synth.studio.models/">https://www.instagram.com/synth.studio.models/</a>这凝聚了许多小时的心血。如果你喜欢，请 👍 点赞、💬 评论，也欢迎 ⚡ 打赏 😉<h3 id="https:ko-fi.comlonecatone">请我喝杯☕<a target="_blank" rel="ugc" href="https://ko-fi.com/lonecatone">https://ko-fi.com/lonecatone</a></h3><hr /><h3 id="lc-vision">LC Vision 🔬📝🎥 — 这个节点包不会浪费你的时间</h3>前言：让支持视觉的 llama.cpp 正常运行起来，通常是一场持续数小时的折腾。Qwen-VL 支持<code>Qwen3VLChatHandler</code> / <code>Qwen25VLChatHandler</code>）从未合并进上游的 <code>llama-cpp-python</code>，只存在于 JamePeng 的 fork 里，以 GitHub Release 的原始 wheel 形式发布，每一个都对应精确的 Python 版本 + CUDA 版本 + 操作系统组合。选错一个，你得到的只是一个莫名其妙的导入错误，而不是任何有用的提示。这个节点包会自动帮你完成这个匹配，在你打开工作流之前就已经搞定。⚠️ <a target="_blank" rel="ugc" href="http://install.py">install.py</a> 会自己搭建环境- 在 ComfyUI-Manager 安装或更新这个节点包时自动运行，你不需要手动执行任何操作。- 首先检查是否已经可以导入支持视觉的 <code>llama_cpp</code>。如果可以，它什么都不做，不会干扰你。- 如果不行，它会直接读取你机器上的 Python 版本、操作系统和 CUDA 版本，拉取 JamePeng/llama-cpp-python 的真实 GitHub Releases 列表，安装与之匹配的那一个 wheel（如果没有精确匹配，会向下取最接近的 CUDA 版本）。- 在做任何改动之前先记录一份 <code>pip freeze</code> 快照，并且只在 <code>pip check</code> 事后报告了真实冲突、且问题确实出在 <code>llama-cpp-python</code> 自身时才会介入，绝不会动其他包的版本锁定。你环境里的其他一切都不会被碰。- 每次更新时重复运行都是安全的。这是设计上的幂等，不是运气。Loader 到底是怎么加载模型的- <code>model_name</code> 会列出 <code>models/LLM</code> 目录下已有的每一对 <code>.gguf</code> 及其匹配的 <code>mmproj</code>，以及一组带有 <code>⬇ Download:</code> 前缀的精选一键下载选项。- 选择一个 <code>⬇ Download:</code> 选项，它会在首次使用时直接从 HuggingFace 下载，并存放到同一个目录下。下次它就会作为普通选项出现在列表里。- 💡 目前精选的模型- Qwen3-VL-4B-abliterated 和 Qwen3-VL-8B-abliterated，各自提供 Q8_0（更小更快）和 f16（完整精度）两个版本，一共四个，都是精心挑选的，而不是把上游整个庞大目录甩给你。- 4B 是视频 / Moviemaker 的首选，为多帧批处理留出更多余量。8B 是图像 / Caption 的首选，保真度更高。- 无论主模型选择哪种量化，mmproj 始终以 f16 下载。视觉质量更依赖 mmproj 的精度，而不是语言模型权重本身。- 模型一旦加载就会保持常驻。只要输入没有变化，ComfyUI 自身的节点缓存机制就会完全跳过 Loader 的重新执行，和模型检查点加载器一样。每次调用 Caption、Enhancer 或 Moviemaker 之间都不用重新支付加载成本。放心用：其他视觉节点包给你的是一份 requirements.txt 和一个让你自己去找对应 wheel 的 wiki 页面。这个节点包把 fork 检测、CUDA 匹配和下载全部内置好了。放下节点，选好模型，直接就能跑。<hr /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/f5090818-66ec-4f1d-b7f6-4fe05fae46e8/original=true/f5090818-66ec-4f1d-b7f6-4fe05fae46e8.jpeg" />"True Nothing is. Permitted Everything is"- Yoda Auditore, Assassin's Wars

Files with SHA256: d9c570fc1c1a4e5d3ed477fd0df44cf27026e93dcc71b3d6c63764debcbe3b3e