报错
112
本机环境专项
含未核实说明
GitHub 与 huggingface.co 直连超时:自定义节点装不上、模型下载中断(本机网络实测)
报错原文(原样)
(本机实测,PowerShell 直接请求) https://github.com/comfyanonymous/ComfyUI/issues/1 => FAIL: The operation has timed out. (12094 ms) https://docs.comfy.org/get_started/introduction => FAIL: The operation has timed out. (12011 ms) https://raw.githubusercontent.com => FAIL: The operation has timed out. https://huggingface.co => FAIL: The operation has timed out. https://hf-mirror.com => 200 OK (602 ms) https://mirrors.aliyun.com/pytorch-wheels/cu130/ => 200 OK https://mirrors.tuna.tsinghua.edu.cn/help/pypi/ => 200 OK https://ghproxy.net/ => 200 OK
检索关键词
The operation has timed out、Failed to connect to github.com、Could not connect to huggingface.co、We couldn't connect to 'https://huggingface.co'、Manager 安装节点失败、模型下载不动
先看这一句
浏览器明明能上网,但 ComfyUI-Manager 里点安装节点一直转圈或失败,模型库下载不动,git clone 超时;控制台里常见连接超时或 SSL 握手失败。
适用版本 网络环境相关问题,与 ComfyUI 版本无关
·
核实日期 2026-09-26
·
状态 可能已过期,正在复核
适用环境:Windows 便携版、秋叶/绘世整合包、全部主流环境
现象与触发场景
常见触发场景:
- ComfyUI-Manager 安装 / 更新自定义节点时卡在下载步骤
- 从 HuggingFace 或 Civitai 下载模型时中断、超时
- 工作流里某个节点要通过 URL 自动拉取模型文件时失败
- pip install 装节点依赖时连不上 pypi 或 GitHub 托管的源码包
根因
- 本机所在网络到 github.com、raw.githubusercontent.com、huggingface.co 的直连不可用(实测均为超时),这三个域名恰好是自定义节点与模型的主要托管地。
- 本机系统里配置过代理 127.0.0.1:7892,但当前 ProxyEnable=0 且该端口没有进程监听,即代理处于关闭状态(本机实测),所以 ComfyUI 走的是直连。
- 注意:只有 HTTP 请求超时的域名受影响,DNS 与国内站点正常,因此这不是断网,而是特定域名的可达性问题。
解决步骤
已按「最可能有效」排序。请从上往下做,每做完一步用「验证」确认,不要一次改五处。
- 先做 30 秒可达性自检,确认到底是哪一层出问题在 PowerShell 里跑这几条,能立刻区分是 DNS 问题、TLS 问题还是纯超时。本机实测结果见本条 error_text。Test-NetConnection github.com -Port 443Test-NetConnection huggingface.co -Port 443Invoke-WebRequest -Uri https://hf-mirror.com -Method Head -TimeoutSec 10 -UseBasicParsing | Select-Object StatusCode
- 给 ComfyUI 挂上 HuggingFace 镜像(对模型下载最有效)HF_ENDPOINT 是 huggingface_hub 官方支持的环境变量,指向镜像后所有 HF 下载都会走镜像。写进启动脚本是最省事的做法,这样以后每次启动都生效,不用每次手动设置。set HF_ENDPOINT=https://hf-mirror.com(在 C:\ComfyUI\启动ComfyUI.bat 里,set "PATH=..." 那一行之后加入:set "HF_ENDPOINT=https://hf-mirror.com")
- 把 pip 切到国内镜像,解决节点依赖装不上清华 TUNA 镜像稳定且支持 pip。配置一次永久生效,比每次加 -i 参数省事。C:\ComfyUI\python\python.exe -m pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simpleC:\ComfyUI\python\python.exe -m pip config set global.trusted-host pypi.tuna.tsinghua.edu.cn
- PyTorch 相关包一律走阿里云 wheel 镜像下载 torch 这类大包时,阿里云镜像比 pypi 和官方源都快,且 cu118 / cu128 / cu130 各版本目录齐全。C:\ComfyUI\python\python.exe -m pip install --no-deps --force-reinstall <包名>.whl --index-url https://mirrors.aliyun.com/pytorch-wheels/cu128/
- 需要读 GitHub 上的单个文件时用 ghproxy 前缀本机实测 ghproxy.net 可以正常代理 raw.githubusercontent.com 的文件(返回 200),但对 GitHub 的网页与 API 请求会返回 403,所以它只适合取原始文件(例如 README、requirements.txt、单个脚本),不适合浏览 issue。https://ghproxy.net/https://raw.githubusercontent.com/<owner>/<repo>/<branch>/<path>
- 要读 ComfyUI 官方源码或配置文件时,走 gitee 镜像的 raw 通道(本机实测最快)gitee 上有 ComfyUI 主仓库的镜像,raw 文件读取实测可用且速度快:本机实测 comfy/cli_args.py、comfy/model_management.py、README.md 均返回 200,耗时 350-500 ms,比 ghproxy 稳定(ghproxy 对 GitHub 网页与 API 会返回 403)。适合快速取官方文件、看目录结构、拿 requirements 清单。⚠ 但实测发现该镜像会滞后于官方:同一时刻 gitee 镜像的 pyproject.toml 显示 version = "0.36.0",而官方仓库(Comfy-Org/ComfyUI)与本机均为 0.37.0。因此用它核对「某个参数在当前版本是否还存在」这类版本相关结论时,必须先确认镜像里的版本号与你本机一致,否则会得出过时结论。需要严格对照官方最新版时,改用 ghproxy 前缀读官方 raw(实测能取到 0.37.0 的文件,见下面的命令)。https://gitee.com/mirrors/ComfyUI/raw/master/comfy/cli_args.pyhttps://gitee.com/mirrors/ComfyUI/raw/master/comfy/model_management.pyhttps://ghproxy.net/https://raw.githubusercontent.com/Comfy-Org/ComfyUI/master/pyproject.toml
- 模型无法自动下载时,改为手动下载后放进 models 目录这是最稳的兜底方案:在能上网的设备或用镜像把文件下载好,再按类型放进 C:\ComfyUI\ComfyUI\models 下对应子目录(checkpoints / loras / vae / clip / unet / controlnet 等)。放完后重启 ComfyUI,节点下拉框里就能看到。
- 确实需要直连 GitHub 时,启用本机已配置好的代理本机系统代理配置里已有 127.0.0.1:7892(ProxyEnable=0 说明当前关闭),启动对应代理软件后再把开关打开即可。如果希望 ComfyUI 也走代理,可以在启动脚本里加环境变量。set HTTP_PROXY=http://127.0.0.1:7892set HTTPS_PROXY=http://127.0.0.1:7892
验证是否修好
在 ComfyUI 自己的解释器里验证是否已生效(这一步很关键,因为它走的是 ComfyUI 内置 Python 而不是系统 Python):执行 C:\ComfyUI\python\python.exe -c "import os;print(os.environ.get('HF_ENDPOINT'))",应打印出 https://hf-mirror.com;然后在 Manager 里重新安装之前失败的节点,观察是否还卡在下载。
补充说明
[本机] 本条全部结论来自本机实测:2026-09-26 用 PowerShell 逐一请求上述域名,github.com / docs.comfy.org / raw.githubusercontent.com / huggingface.co 均超时,hf-mirror.com / 阿里云镜像 / 清华镜像 / ghproxy.net 均返回 200。 [本机] 代理状态来自注册表 HKCU\Software\Microsoft\Windows\CurrentVersion\Internet Settings(ProxyEnable=0,ProxyServer=127.0.0.1:7892),且该端口当前无监听进程。 [未核实] 各镜像站的长期可用性与限速策略会变化,使用前请重新测试;本机实测结果不代表未来状态。 [本机] 补充实测:gitee 镜像 raw 通道可用(三个官方源码文件均 200,350-500 ms),已加入解决步骤。
来源
按可信度排列:官方 issue / 官方文档 > 节点仓库 issue > 社区帖 > 中文社区文章。
链接以纯文本给出(本站不做站外跳转),需要核对时请自行复制到浏览器打开。
- [161] (镜像站(本机实测)) HF-Mirror(HuggingFace 国内镜像站,本机实测 HTTP 200) https://hf-mirror.com/
- [258] (镜像站帮助页(本机实测)) 清华大学开源软件镜像站 —— PyPI 镜像使用帮助(本机实测 HTTP 200) https://mirrors.tuna.tsinghua.edu.cn/help/pypi/
- [257] (镜像站(本机实测)) 阿里云 PyTorch wheel 镜像(本机实测 HTTP 200) https://mirrors.aliyun.com/pytorch-wheels/cu130/
- [259] (代理服务(本机实测)) ghproxy —— GitHub 文件代理(本机实测 HTTP 200,仅对 raw 文件有效) https://ghproxy.net/
- 附录 A 本机环境基线(实测)
- 以下数据来自本机实际探测,用于判断报错是否与版本组合相关。别人的报错解法不一定适用于本机,先核对版本再动手。
- 手册生成日期
- 见封面
- ComfyUI 版本
- 0.37.0(读取 C:\ComfyUI\ComfyUI\comfyui_version.py 得到)
- 安装形态
- Windows 便携式自建部署,根目录 C:\ComfyUI(内含 ComfyUI\、python\、ffmpeg\ 三个子目录)
- Python 解释器
- 3.12.9(C:\ComfyUI\python\python.exe,MSC v.1942 64 bit)
- PyTorch
- 2.11.0+cu128
- CUDA 运行时
- 12.8
- 显卡
- NVIDIA GeForce RTX 5060 Ti(Blackwell 架构,计算能力 sm_120)
- 显存 / 内存
- Total VRAM 16311 MB(16 GB) / Total RAM 32546 MB(32 GB)
- 已装自定义节点
- 仅 ComfyUI-Manager(custom_nodes 目录实测)
- 启动方式
- C:\ComfyUI\启动ComfyUI.bat → python\python.exe -s main.py --auto-launch --listen 127.0.0.1
- 自检脚本
- C:\ComfyUI\体检.bat → check.ps1(安装时生成)
- VRAM 模式
- NORMAL_VRAM
- Attention 实现
- Using pytorch attention(未使用 xformers / SageAttention)
- 权重卸载
- async weight offloading with 2 streams;pinned memory 13018.0
- models 常用子目录(实测)
- checkpoints / loras / vae / clip / text_encoders / clip_vision / unet / diffusion_models / controlnet / embeddings / upscale_models / vae_approx(本机 models 目录下共 25 个子目录,新旧命名并存;放错子目录是模型加载报错的第一大原因)
- ffmpeg
- C:\ComfyUI\ffmpeg\ffmpeg.exe 已存在,启动脚本会把该目录加入 PATH(视频节点报 ffmpeg not found 时可先确认这一项)
- extra_model_paths.yaml
- 本机不存在(未配置额外模型路径;若要复用另一个 ComfyUI 安装的模型库,需要自己创建这个文件)
- 实测待办①
- 启动日志有 WARNING:You need pytorch with cu130 or higher to use optimized CUDA operations. / If you are on nvidia 20 series and above it is required that you update your pytorch to cu130 or higher.
- 实测待办②
- comfy_kitchen 后端自检结果:cuda: available=True 但 disabled=True(因 torch 为 cu128 而非 cu130);eager 可用;hip 不可用;triton 不可用
- 实测待办③
- triton 未安装(日志原文:ImportError: No module named 'triton'),所以依赖 triton 的加速节点(SageAttention 等)在本机不可用
- ★ 路径差异必读:手册正文里凡是出现 python_embeded\python.exe 的命令,都是按 ComfyUI 官方 Windows 便携版的目录结构写的(官方包把解释器放在 python_embeded 子目录)。本机是自建部署,没有 python_embeded 目录,解释器在 C:\ComfyUI\python\python.exe。遇到这类命令,请把 python_embeded\python.exe 整体替换成 C:\ComfyUI\python\python.exe 再执行。第 11 类「本机实测专项」里的命令已经全部按本机真实路径写好,可以直接复制执行。
- 本机启动脚本 C:\ComfyUI\启动ComfyUI.bat 会先把 C:\ComfyUI\ffmpeg、C:\ComfyUI\python、C:\ComfyUI\python\Scripts 加进 PATH,再进 C:\ComfyUI\ComfyUI 目录执行 python\python.exe -s main.py --auto-launch --listen 127.0.0.1。所以在本机,只要是用这个 bat 启动的 ComfyUI,pip 必须用 C:\ComfyUI\python\python.exe -m pip 调用,直接敲 pip 有可能装到别的 Python 里。
- 以上全部为 [本地] 实测数据,来源:ComfyUI 启动日志 C:\ComfyUI\ComfyUI\user\comfyui_8188.log、C:\ComfyUI\安装日志.txt、C:\ComfyUI\启动ComfyUI.bat、C:\ComfyUI\体检.bat,以及用 C:\ComfyUI\python\python.exe 直接查询 torch 得到的结果。
- 本机提示「需升级到 cu130」这一条,属于 ComfyUI 新版本对 Blackwell(50 系)显卡的优化路径要求:当前 cu128 环境能正常出图,但部分新版优化算子(comfy_kitchen 的 CUDA 后端)被自动禁用。是否升级请自行权衡——升级 torch/CUDA 属于高风险操作,务必先备份 C:\ComfyUI\python 目录(或至少导出 pip freeze 清单)。第 11 类有完整的前置条件核实结果与回滚命令。
- 本机依赖由自建安装流程部署,安装日志中引用的依赖清单地址为 E:\comfyui网站包\comfyui依赖安装\09-清单\ 下的 requirements-*.txt。改动依赖前建议先回到这套清单,而不是随手 pip install。
- 本手册正文中的报错条目来自公开网络来源(官方 issue、官方文档、节点仓库 issue、社区帖),与上表本机数据分开标注:正文说「全版本通用」是指网络上多个来源一致,不代表已在本机复现。
- 附录 B 通用排障流程(不知道从哪下手时用)
- 1. 确认它到底起没起来
- 看两个地方:黑窗口里最后一行是不是 Starting server / To see the GUI go to: http://127.0.0.1:8188;浏览器访问 http://127.0.0.1:8188/system_stats 是否能返回 JSON。后者能返回,说明后端没问题,问题在前端或浏览器。
- 2. 分清三类故障
- 启动就挂(进程根本没起来)→ 查安装与环境类;起来了但一点就报错(某个节点执行失败)→ 查运行时/节点类;全程很慢或不报错但出黑图 → 查显存、精度、模型类。
- 3. 读 Traceback 的顺序
- 从下往上读:最后一行是错误类型与消息(这是检索钥匙);往上找最后一个出现在 C:\ComfyUI\ComfyUI\ 或 custom_nodes\ 里的文件路径,那一行大概率就是出错节点。
- 4. 用最小可复现集定位
- 把工作流删到只剩出错节点及其上下游,逐个断开重连,定位到具体是哪两个节点组合出问题。这比读日志快。
- 5. 二分禁用自定义节点
- 把 custom_nodes 里一半文件夹改名加 .bak,重启看是否恢复,再对另一半重复。三到四轮就能锁定是哪个节点包搞坏了整个启动。
- 6. 恢复出厂环境
- 把所有自定义节点移出 custom_nodes,用官方默认工作流测试。默认工作流出图正常,说明 ComfyUI 本体没问题,问题在节点或模型。
- 7. 保留证据再求助
- 提问时给三样东西:ComfyUI 启动时打印的版本信息整段、完整 Traceback、你的显卡与驱动版本(nvidia-smi 截图)。缺任何一样,对方都只能猜。
- 附录 C 来源清单
- 本手册共引用 259 个来源页面。每条报错正文里的 [n] 编号对应下表。
- 编号
- 来源页面
- 链接
- 1
- (官方仓库文件)ComfyUI/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat
- https://github.com/Comfy-Org/ComfyUI/blob/master/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat
- 2
- (官方论坛)Wan 2.2 workflows suddenly failing halfway
- https://forum.comfy.org/t/wan-2-2-workflows-suddenly-failing-halfway/4252
- 3
- (官方文档)System Requirements - ComfyUI
- https://docs.comfy.org/installation/system_requirements
- 4
- (官方 issue)[Errno 48] error while attempting to bind on address ('127.0.0.1', 8188): address already in use
- https://github.com/Comfy-Org/ComfyUI/issues/1579
- 5
- (社区)ComfyUI-Stuck-Process-Killer: A simple script to kill stuck ComfyUI processes that are blocking port 8188
- https://github.com/Zuowa111/ComfyUI-Stuck-Process-Killer
- 6
- (官方文档)How to Troubleshoot and Solve ComfyUI Issues - ComfyUI
- https://docs.comfy.org/troubleshooting/overview
- 7
- (官方论坛)Comfy won't start - No Available Ports
- https://forum.comfy.org/t/comfy-wont-start-no-available-ports/1225
- 8
- (官方论坛)ComfyUI Desktop 0.7.2 could not start
- https://forum.comfy.org/t/comfyui-desktop-0-7-2-could-not-start/4208
- 9
- (官方文档)Install ComfyUI Manually (Windows, macOS, Linux)
- https://docs.comfy.org/installation/manual_install
- 10
- (官方 issue)ComfyUI Portable cannot find a dll that is clearly there
- https://github.com/Comfy-Org/ComfyUI/issues/10535
- 11
- (中文社区)Windows下ComfyUI启动失败,提示“Python未找到”如何解决?
- https://ask.csdn.net/questions/9460791
- 12
- (官方论坛)Was not able to start ComfyUI
- https://forum.comfy.org/t/was-not-able-to-start-comfyui/38
- 13
- (官方论坛)Can't install ComfyUi, probably because of the Polish letter Ł in the file path
- https://forum.comfy.org/t/cant-install-comfyui-probably-because-of-the-polish-letter-l-in-the-file-path/1248
- 14
- (官方论坛)Error: Failed to create virtual environment: exit code -999
- https://forum.comfy.org/t/error-failed-to-create-virtual-environment-exit-code-999-with-the-issue-pretty-obvious-at-the-end/390
- 15
- (官方论坛)Can't install ComfyUI 0.3.15
- https://forum.comfy.org/t/cant-install-comfyui-0-3-15/483
- 16
- (官方论坛)Unable to start ComfyUI desktop version v0.4.3 on M4Mac(Sequoia 15.2)
- https://forum.comfy.org/t/unable-to-start-comfyui-desktop-version-v0-4-3-on-m4mac-sequoia-15-2/926
- 17
- (官方论坛)Recreating venv installs non-functioning version of torch without the proper backend
- https://forum.comfy.org/t/recreating-venv-installs-non-functioning-version-of-torch-without-the-proper-backend/4314
- 18
- (官方文档)ComfyUI README — Troubleshooting(Torch not compiled with CUDA enabled)
- https://github.com/Comfy-Org/ComfyUI/blob/master/README.md
- 19
- (官方论坛)Issues after Updating Desktop Version (code 3221225477)
- https://forum.comfy.org/t/issues-after-updating-desktop-version-code-3221225477/4133
- 20
- (官方仓库文件)ComfyUI pyproject.toml(requires-python = ">=3.10")
- https://github.com/comfyanonymous/ComfyUI/blob/master/pyproject.toml
- 21
- (中文社区)ComfyUI秋叶整合包下载后无法启动,提示Python版本错误怎么办?
- https://ask.csdn.net/questions/9298946
- 22
- (官方论坛)Unable to start ComfyUI Desktop v0.4.72
- https://forum.comfy.org/t/unable-to-start-comfyui-desktop-v0-4-72/3800
- 23
- (官方文档)Comfy Desktop FAQ - ComfyUI
- https://docs.comfy.org/installation/desktop/faq
- 24
- (官方文档)如何更新 ComfyUI - ComfyUI(依赖更新故障排除)
- https://docs.comfy.org/zh/installation/update_comfyui
- 25
- (官方 issue)accidently ran "update_comfyui_and_python_dependencies.bat" and broke ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/7415
- 26
- (官方 issue)AssertionError: Torch not compiled with CUDA enabled after running update_comfyui_and_python_depende
- https://github.com/Comfy-Org/ComfyUI/issues/11575
- 27
- (官方 issue)[Announcement] The frontend will no longer be shipped in the main ComfyUI repo, it will be a separat
- https://github.com/Comfy-Org/ComfyUI/issues/7025
- 28
- (官方论坛)Error ComfyUI refuses to run - Error: no available ports found between 8000 and 9000
- https://forum.comfy.org/t/error-comfyui-refuses-to-run-error-no-available-ports-found-between-8000-and-9000/3798
- 29
- (官方论坛)Recommended approach to enable local network access doesn't work
- https://forum.comfy.org/t/recommended-approach-to-enable-local-network-access-doesnt-work/1747
- 30
- (官方论坛)Cant run comfyui on lan network
- https://forum.comfy.org/t/cant-run-comfyui-on-lan-network/2668
- 31
- (官方论坛)Starting ComfyUI server... -> Unable to start ComfyUI Desktop - Comfy v0.8.4
- https://forum.comfy.org/t/starting-comfyui-server-unable-to-start-comfyui-desktop-comfy-v0-8-4/4282
- 32
- (官方论坛)Malwarebytes Quarantines Python File
- https://forum.comfy.org/t/malwarebytes-quarantines-python-file/111
- 33
- (官方论坛)Unable to start ComfyUI Desktop V0.8.23
- https://forum.comfy.org/t/unable-to-start-comfyui-desktop-v0-8-23/4334
- 34
- (官方 issue)pls help ive been struggling for days error after error getting ran out of memory error · Issue #127
- https://github.com/Comfy-Org/ComfyUI/issues/12713
- 35
- (官方文档)ComfyUI/comfy/cli_args.py(参数与 help 原文)
- https://github.com/Comfy-Org/ComfyUI/blob/master/comfy/cli_args.py
- 36
- (官方文档)ComfyUI/comfy/model_management.py(EXTRA_RESERVED_VRAM 等)
- https://github.com/Comfy-Org/ComfyUI/blob/master/comfy/model_management.py
- 37
- (官方 issue)comfyui image to video on 6GB GPU, torch.OutOfMemoryError · Issue #14468
- https://github.com/Comfy-Org/ComfyUI/issues/14468
- 38
- (官方 issue)CLIPTextEncode Allocation on device ... TIPS: batch_size · Issue #9513
- https://github.com/Comfy-Org/ComfyUI/issues/9513
- 39
- (官方 issue)WanVideoSampler Allocation on device ... TIPS: batch_size · Issue #10194
- https://github.com/Comfy-Org/ComfyUI/issues/10194
- 40
- (官方 issue)torch.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. · Issue #13954
- https://github.com/Comfy-Org/ComfyUI/issues/13954
- 41
- (官方 issue)VAE decoding on AMD works first time only before switching to tiled or black image · Issue #9547
- https://github.com/Comfy-Org/ComfyUI/issues/9547
- 42
- (官方 issue)Free other models' memory before retrying handles_tiling VAE decode (PR #15456)
- https://github.com/Comfy-Org/ComfyUI/pull/15456
- 43
- (官方文档)VAEDecodeTiled - ComfyUI Built-in Node Documentation
- https://docs.comfy.org/zh/built-in-nodes/VAEDecodeTiled
- 44
- (官方 issue)Flux2: torch.OutOfMemoryError: Allocation on device · Issue #10891
- https://github.com/Comfy-Org/ComfyUI/issues/10891
- 45
- (官方 issue)What should I do. (The paging file is too small, os error 1455) · Issue #13122
- https://github.com/Comfy-Org/ComfyUI/issues/13122
- 46
- (官方 issue)Flux 2 Native workflow got error! UNETLoader OSError os error 1455 · Issue #10889
- https://github.com/Comfy-Org/ComfyUI/issues/10889
- 47
- (社区)ComfyUI 'Paging File Is Too Small' (os error 1455)? Fix Windows Virtual Memory for Local AI
- https://runaihome.com/blog/comfyui-paging-file-too-small-os-error-1455-fix-2026/
- 48
- (中文社区)ComfyUI项目运行中"OS Error 1455"分页文件不足问题的解决方案
- https://blog.gitcode.com/3300ada251c18ade5a89230b223ed06b.html
- 49
- (中文社区)comfyUI 避免GPU使用共享内存的方法
- https://www.cnblogs.com/zxdplay/p/19461876
- 50
- (官方 issue)(Windows) Memory Compression is not freed after a run. · Issue #10338
- https://github.com/Comfy-Org/ComfyUI/issues/10338
- 51
- (官方 issue)torch.cuda.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. (out of memory) · I
- https://github.com/Comfy-Org/ComfyUI/issues/1918
- 52
- (官方 issue)out of memory error but I still have free memory · Issue #9121
- https://github.com/Comfy-Org/ComfyUI/issues/9121
- 53
- (官方 issue)Experiencing VRAM OOM with the same workflow after upgrading to ComfyUI 0.27 · Issue #14789
- https://github.com/Comfy-Org/ComfyUI/issues/14789
- 54
- (官方 issue)Strange VRAM consumption (SDXL barely fits in 8GB VRAM) · Issue #2855
- https://github.com/Comfy-Org/ComfyUI/issues/2855
- 55
- (官方 issue)Current memory manager is a garbage · Issue #15443
- https://github.com/Comfy-Org/ComfyUI/issues/15443
- 56
- (官方 issue)update default command-line arguments for run_amd_gpu.bat · Issue #11624
- https://github.com/Comfy-Org/ComfyUI/issues/11624
- 57
- (官方 issue)Dynamic VRAM prevents native FP8 compute for weight_dtype fp8_e4m3fn_fast · Issue #16551
- https://github.com/Comfy-Org/ComfyUI/issues/16551
- 58
- (官方文档)Performance Optimization - ComfyUI
- https://mintlify.wiki/Comfy-Org/ComfyUI/guides/performance-optimization
- 59
- (官方 issue)Fix/tweak pinned memory accounting (PR #13221)
- https://github.com/Comfy-Org/ComfyUI/pull/13221
- 60
- (官方 issue)Memory is not release after each generation · Issue #10475
- https://github.com/Comfy-Org/ComfyUI/issues/10475
- 61
- (官方 issue)Cumulative system RAM leak requiring full PC restart with Wan2.2 GGUF · Issue #11775
- https://github.com/Comfy-Org/ComfyUI/issues/11775
- 62
- (节点仓库 issue)ComfyUI-Force-Cleanup: Force ComfyUI to unload models and clear execution/device caches after each
- https://github.com/3289004205/ComfyUI-Force-Cleanup
- 63
- (节点仓库 issue)ComfyUI-PipelineBarrier: Force-flush GPU caches between ComfyUI pipeline stages to prevent OOM kil
- https://github.com/brosequist/ComfyUI-PipelineBarrier
- 64
- (社区)GPU memory contention with other heavy CUDA apps
- https://github.com/Rimagination/h3lite/blob/main/references/gpu-contention.md
- 65
- (节点仓库 issue)IPAdapterInsightFaceLoader No module named 'insightface' · Issue #770 · cubiq/ComfyUI_IPAdapter_pl
- https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/770
- 66
- (社区)Prebuilt Windows wheel for insightface==0.7.3 (Python 3.12, cp312, amd64)
- https://github.com/skippzz/insightface-0.7.3-cp312-cp312-win_amd64
- 67
- (节点仓库 issue)i am getting error when try to run install.bat · Issue #154 · Gourieff/ComfyUI-ReActor
- https://github.com/Gourieff/ComfyUI-ReActor/issues/154
- 68
- (官方文档)自定义节点问题排查 / Troubleshooting Custom Node Issues
- https://docs.comfy.org/troubleshooting/custom-node-issues
- 69
- (官方 issue)comfy failing to import any custom node after update with comfy-cli · Issue #1730
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/1730
- 70
- (官方 issue)Custom nodes no longer being imported when Comfy starts · Issue #7554
- https://github.com/Comfy-Org/ComfyUI/issues/7554
- 71
- (节点仓库 issue)ComfyUI - Control Net Aux - ModuleNotFoundError: No module named 'cv2' · Issue #153
- https://github.com/Fannovel16/comfyui_controlnet_aux/issues/153
- 72
- (节点仓库 issue)IMPORT FAILED: ModuleNotFoundError: No module named 'cv2' · Issue #674
- https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/674
- 73
- (节点仓库 issue)(IMPORT FAILED): C:\ComfyUI\custom_nodes\ComfyUI-WD14-Tagger · Issue #65
- https://github.com/pythongosssss/ComfyUI-WD14-Tagger/issues/65
- 74
- (官方 issue)Importing onnxruntime_pybind11_state fails in Windows Python 3.10 · Issue #16116 · microsoft/onnxrun
- https://github.com/microsoft/onnxruntime/issues/16116
- 75
- (节点仓库 issue)cannot import name 'cached_download' from 'huggingface_hub' - Magic Clothing Import Failed - Issue
- https://github.com/frankchieng/ComfyUI_MagicClothing/issues/80
- 76
- (社区)ImportError: cannot import name 'cached_download' from 'huggingface_hub'
- https://stackoverflow.com/questions/79374322/importerror-cannot-import-name-cached-download-from-huggingface-hub
- 77
- (节点仓库 issue)Failed import · Issue #76 · kijai/ComfyUI-CogVideoXWrapper
- https://github.com/kijai/ComfyUI-CogVideoXWrapper/issues/76
- 78
- (节点仓库 issue)Missing nodes - Tried many ways including clone Github and run python requirements · Issue #15
- https://github.com/bytedance/ComfyUI_InfiniteYou/issues/15
- 79
- (官方文档)使用 ComfyUI-Manager 进行自定义节点管理(旧版 UI)
- https://docs.comfy.org/manager/legacy-ui
- 80
- (官方文档)ComfyUI Manager Troubleshooting
- https://docs.comfy.org/manager/troubleshooting
- 81
- (官方文档)ComfyUI-Manager 配置(config.ini 选项与安全级别)
- https://docs.comfy.org/manager/configuration
- 82
- (节点仓库 issue)Error building insightface? · Issue #14 · ZHO-ZHO-ZHO/ComfyUI-InstantID
- https://github.com/ZHO-ZHO-ZHO/ComfyUI-InstantID/issues/14
- 83
- (社区)Building wheel for insightface (pyproject.toml): finished with status 'error' · Issue #2696
- https://github.com/deepinsight/insightface/issues/2696
- 84
- (节点仓库 issue)[Windows/Python 3.12] Critical Installation Failure: GroundingDINO Dependency · Issue #27
- https://github.com/judian17/ComfyUI-SDPose-OOD/issues/27
- 85
- (节点仓库 issue)This node is impossible to install. · Issue #216 · Gourieff/ComfyUI-ReActor
- https://github.com/Gourieff/ComfyUI-ReActor/issues/216
- 86
- (节点仓库 issue)problem to install the Install the required Python packages · Issue #39
- https://github.com/1038lab/ComfyUI-OmniGen/issues/39
- 87
- (官方 issue)Custom node dependencies not installed when imported on startup · Issue #6035
- https://github.com/Comfy-Org/ComfyUI/issues/6035
- 88
- (官方文档)cm-cli: ComfyUI-Manager CLI(restore-dependencies / fix 命令)
- https://raw.githubusercontent.com/Comfy-Org/ComfyUI-Manager/main/docs/en/cm-cli.md
- 89
- (社区)Safetensors HeaderTooLarge Error? Your Model File Isn't a Model — the 60-Second Diagnosis (2026)
- https://runaihome.com/blog/safetensors-headertoolarge-error-fix-2026/
- 90
- (官方 issue)CLIPVisionLoader Error while deserializing header: HeaderTooLarge · Issue #5797 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/5797
- 91
- (节点仓库 issue)safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLarge · Issue #17 · d
- https://github.com/diontimmer/ComfyUI-Vextra-Nodes/issues/17
- 92
- (官方 issue)Header too large error · Issue #7510 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/7510
- 93
- (官方文档)如何排查和解决 ComfyUI 中模型相关的问题 - ComfyUI 官方文档
- https://docs.comfy.org/zh/troubleshooting/model-issues
- 94
- (官方文档)Error while deserializing header: incomplete metadata, file not fully covered - Custom Nodes - ComfyUI 官
- https://forum.comfy.org/t/error-while-deserializing-header-incomplete-metadata-file-not-fully-covered/3797
- 95
- (节点仓库 issue)safetensors_rust.SafetensorError: Error while deserializing header: MetadataIncompleteBuffer · Iss
- https://github.com/kijai/ComfyUI-CogVideoXWrapper/issues/99
- 96
- (官方 issue)FlashVsr. LoadWanVideoT5TextEncoder Error while deserializing header: incomplete metadata, file not
- https://github.com/Comfy-Org/ComfyUI/issues/10675
- 97
- (官方文档)模型 - ComfyUI 官方文档(目录结构与 extra_model_paths.yaml)
- https://docs.comfy.org/zh/basic-concepts/models
- 98
- (官方文档)Model Paths Configuration - ComfyUI
- https://mintlify.wiki/Comfy-Org/ComfyUI/configuration/model-paths
- 99
- (社区)Value not in list (model not found) — ComfyUI Error Fix
- https://www.numonic.ai/integrations/comfyui-errors/value-not-in-list
- 100
- (社区)250928 low noise:lora key not loaded issue - lora key not loaded: blocks.0.cross_attn... · lightx2v/Wan2.2
- https://huggingface.co/lightx2v/Wan2.2-Lightning/discussions/42
- 101
- (官方 issue)Z-Image Turbo LoKR: "lora key not loaded" and weights ignored (LoRA works) · Issue #10973 · Comfy-Or
- https://github.com/Comfy-Org/ComfyUI/issues/10973
- 102
- (节点仓库 issue)When using Z-Image LoRA with ComfyUI, the message "lora key not loaded" appears · Issue #538 · ost
- https://github.com/ostris/ai-toolkit/issues/538
- 103
- (社区)LoRA loading errors with Wan2_2-Animate-14B_fp8_scaled_e4m3fn_KJ_v2.safetensors · Kijai/WanVideo_comfy_fp8
- https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/discussions/24
- 104
- (节点仓库 issue)IPAdapter Unified Loader throw KeyError: 'clipvision' and my solution · Issue #627 · cubiq/ComfyUI
- https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/627
- 105
- (中文社区)Comfyui运行工作流提示IPAdapter加载器ClipVision model not found.
- https://www.51zxw.net/TechArticleDetails.aspx?zid=129&Id=2754
- 106
- (社区)ComfyUI Quant Load Errors: GGUF, fp8, NVFP4, INT8
- https://localaimaster.com/blog/comfyui-quantization-format-errors
- 107
- (节点仓库 issue)Unable to load UnetLoaderGGUF · Issue #117 · city96/ComfyUI-GGUF
- https://github.com/city96/ComfyUI-GGUF/issues/117
- 108
- (节点仓库 issue)Subdirectories inside \models\unet, are not read · Issue #130 · city96/ComfyUI-GGUF
- https://github.com/city96/ComfyUI-GGUF/issues/130
- 109
- (节点仓库 issue)ComfyUI-GGUF/tools/llama.cpp/ggml/src/ggml.c:22112: GGML_ASSERT(info->ne[i] > 0) failed is back! ·
- https://github.com/city96/ComfyUI-GGUF/issues/345
- 110
- (官方 issue)Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype. · Is
- https://github.com/Comfy-Org/ComfyUI/issues/6859
- 111
- (节点仓库 issue)baddbmm_cuda not implemented for 'Float8_e4m3fn' – FP8 error · Issue #721 · kijai/ComfyUI-WanVideo
- https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/721
- 112
- (官方 issue)ComfyUI crashes when loading umt5_xxl_fp8_e4m3fn_scaled.safetensors · Issue #10471 · Comfy-Org/Comfy
- https://github.com/Comfy-Org/ComfyUI/issues/10471
- 113
- (中文社区)SD-Trainer + ComfyUI 部署指南(AutoDL平台)— Flux 四件套目录结构
- https://www.cnblogs.com/HiroMuraki/p/19497023
- 114
- (官方 issue)Flux 2 Klein models CLIPLoader errors · Issue #11951 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/11951
- 115
- (官方文档)How to Troubleshoot and Solve ComfyUI Model Issues - ComfyUI 官方文档
- https://docs.comfy.org/troubleshooting/model-issues
- 116
- (社区)Fix LTX-2 Errors in ComfyUI: OOM, Black Frames & Flicker Solutions
- https://wavespeed.ai/blog/posts/blog-fix-ltx-2-comfyui-errors/
- 117
- (节点仓库 issue)Getting a "Ran out of input" error from one of the dependencies when comfyui starts with segment a
- https://github.com/storyicon/comfyui_segment_anything/issues/47
- 118
- (官方 issue)Make it explicit that PyTorchStreamReader miniz errors likely indicate corrupt checkpoint · PR #1702
- https://github.com/pytorch/pytorch/pull/170244
- 119
- (节点仓库 issue)An error occurs when executing the FILM_VFI node · Issue #90 · Fannovel16/ComfyUI-Frame-Interpolat
- https://github.com/Fannovel16/ComfyUI-Frame-Interpolation/issues/90
- 120
- (官方 issue)Excessive Memory Overcommit on High-RAM Systems · Issue #8298 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/8298
- 121
- (官方 issue)Recent memory changes cause system freeze · Issue #10723 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/10723
- 122
- (官方 issue)The size of tensor a (81) must match the size of tensor b (16) at non-singleton dimension 1
- https://github.com/Comfy-Org/ComfyUI/issues/7444
- 123
- (官方 issue)Sizes of tensors must match except in dimension 1. Expected size 56 but got size 98 for tensor numbe
- https://github.com/Comfy-Org/ComfyUI/issues/7747
- 124
- (节点仓库 issue)Mismatched tensor size
- https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues/15
- 125
- (节点仓库 issue)Error: mat1 and mat2 shapes cannot be multiplied (1x2048 and 768x3072)
- https://github.com/kijai/ComfyUI-HunyuanVideoWrapper/issues/469
- 126
- (节点仓库 issue)"expected scalar type Half but found Float" in motion_module_ad.py
- https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved/issues/493
- 127
- (节点仓库 issue)RuntimeError: Input type (torch.FloatTensor) and weight type (torch.HalfTensor) should be the same
- https://github.com/laksjdjf/IPAdapter-ComfyUI/issues/17
- 128
- (官方 issue)KSampler - Expected all tensors to be on the same device, but found at least two devices, cpu and cu
- https://github.com/Comfy-Org/ComfyUI/issues/8417
- 129
- (节点仓库 issue)Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!
- https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/1146
- 130
- (官方 issue)'NoneType' object has no attribute 'shape' Issue persists after #1836 fix
- https://github.com/Comfy-Org/ComfyUI/issues/3033
- 131
- (节点仓库 issue)error : 'NoneType' object has no attribute 'shape'
- https://github.com/ltdrdata/ComfyUI-Inspire-Pack/issues/138
- 132
- (社区)What to do when an error occurs
- https://comfyui.nomadoor.net/en/notes/error-handling/
- 133
- (节点仓库 issue)Could not run 'torchvision::nms' with arguments from the 'CUDA' backend.
- https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues/915
- 134
- (社区)Why can only resolutions that are multiples of 8 be generated?
- https://comfyui.nomadoor.net/en/notes/why-multiple-of-8/
- 135
- (社区)Getting "Sizes of tensors must match" error when using ComfyUI WanVideoWrapper (wan2.2) to generate video
- https://stackoverflow.com/questions/79778716/getting-sizes-of-tensors-must-match-error-when-using-comfyui-wanvideowrapper
- 136
- (官方文档)Common Issues and Solutions - ComfyUI
- https://mintlify.wiki/Comfy-Org/ComfyUI/guides/common-issues
- 137
- (官方 issue)AssertionError: Torch not compiled with CUDA enabled on 5060ti · Issue #7826 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/7826
- 138
- (社区)ComfyUI 'Torch not compiled with CUDA enabled'? Every Fix That Works on Windows, Linux, and Mac
- https://dev.to/jovan_chan_9500711396d4e6/comfyui-torch-not-compiled-with-cuda-enabled-every-fix-that-works-on-windows-linux-and-mac-556m
- 139
- (官方文档)How to get ComfyUI running on your new Nvidia 5090 or 5080. - ComfyUI Newsletter
- https://blog.comfy.org/p/how-to-get-comfyui-running-on-your
- 140
- (社区)Pytorch support for sm120 - PyTorch Forums
- https://discuss.pytorch.org/t/pytorch-support-for-sm120/216099
- 141
- (官方 issue)UserWarning: NVIDIA GeForce RTX 5070 Ti with CUDA capability sm_120 is not compatible with the curre
- https://github.com/Comfy-Org/ComfyUI/issues/7127
- 142
- (官方 issue)ComfyUI fails to start with a RuntimeError due to a CUDA version mismatch between PyTorch and TorchA
- https://github.com/Comfy-Org/ComfyUI/issues/14384
- 143
- (社区)[Bug]: Detected that PyTorch and torchvision were compiled with different CUDA versions. PyTorch has CUDA
- https://github.com/Comfy-Org/ComfyUI/issues/10331
- 144
- (社区)RuntimeError: operator torchvision::nms does not exist - vision - PyTorch Forums
- https://discuss.pytorch.org/t/runtimeerror-operator-torchvision-nms-does-not-exist/192829
- 145
- (社区)How do I fix "Torch not compiled with CUDA enabled after using Manager to install a custom node" in ComfyU
- https://forum.smartart.live/topic/521-how-do-i-fix-torch-not-compiled-with-cuda-enabled-after-using-manager-to-install-a-custom-node-in-comfyui/
- 146
- (社区)RTX 5090 survival guide: sm_120, CUDA 12 and 13 side by side, and the xformers trap
- https://dev.to/madcoolseed/rtx-5090-survival-guide-sm120-cuda-12-and-13-side-by-side-and-the-xformers-trap-25nh
- 147
- (社区)ComfyUI on Windows: fixing the CUDA version-mismatch mess - Big Iron
- https://www.bigiron.cc/guides/comfyui-on-windows-cuda-version-mismatch-troubleshooting
- 148
- (节点仓库 issue)FP8 QK kernels unstable on Ada (sm89, RTX 4060 Ti) — BF16/FP8 path causes launch failures, FP16 fa
- https://github.com/thu-ml/SageAttention/issues/309
- 149
- (官方 issue)Regression: RTX 5090 + WAN 2.2 causes cudaErrorIllegalAddress / GPU driver crash after ComfyUI 0.3.6
- https://github.com/Comfy-Org/ComfyUI/issues/14714
- 150
- (社区)【StableDiffusion】CUDAエラーに対するトラブルシューティング【RTX50シリーズ】
- https://elirlab.com/sd-error-cuda-20250914/
- 151
- (官方文档)ComfyUI/README.md(Windows Portable / Troubleshooting)
- https://raw.githubusercontent.com/Comfy-Org/ComfyUI/master/README.md
- 152
- (社区)Linux - PopOs - Torch complains NVIDIA driver is too old, when I have a driver that was released last week
- https://github.com/lllyasviel/Fooocus/issues/2391
- 153
- (节点仓库 issue)Triton fork for Windows support(triton-windows README)
- https://raw.githubusercontent.com/woct0rdho/triton-windows/master/README.md
- 154
- (节点仓库 issue)SageAttention fork for Windows wheels and easy installation
- https://raw.githubusercontent.com/woct0rdho/SageAttention/master/README.md
- 155
- (社区)How do I fix "ModuleNotFoundError: No module named 'triton'" in ComfyUI?
- https://forum.smartart.live/topic/526-how-do-i-fix-modulenotfounderror-no-module-named-triton-in-comfyui/
- 156
- (官方 issue)running update with python dependencies has update numpy to vrs 2 and broken 50% of my custom nodes
- https://github.com/Comfy-Org/ComfyUI/issues/9156
- 157
- (节点仓库 issue)ImportError: numpy.core.multiarray --- Numpy 2.0.0 · Issue #995 · opencv/opencv-python
- https://github.com/opencv/opencv-python/issues/995
- 158
- (节点仓库 issue)`onnxruntime` and `onnxruntime-gpu` can not co-exist and support CUDAExecutionProvider · Issue #26
- https://github.com/microsoft/onnxruntime/issues/26996
- 159
- (节点仓库 issue)[IMPORTANT] Check your Reactor uses GPU -- not CPU!! · Gourieff/ComfyUI-ReActor · Discussion #169
- https://github.com/Gourieff/ComfyUI-ReActor/discussions/169
- 160
- (官方 issue)Cannot connect to host huggingface.co:443 ssl:default [信号灯超时时间已到] · Issue #5184 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/5184
- 161
- (官方文档)🤗 HF-Mirror — 如何使用HF-Mirror
- https://hf-mirror.com/
- 162
- (官方文档)ComfyUI Manager 故障排除 - ComfyUI
- https://docs.comfy.org/zh/manager/troubleshooting
- 163
- (官方 issue)MimicMotionGetPoses · Issue #5613 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/5613
- 164
- (官方 issue)WD14Tagger|pysssss Cannot connect to host cdn-lfs.hf.co:443 ssl:True [SSLCertVerificationError: (1,
- https://github.com/Comfy-Org/ComfyUI/issues/5096
- 165
- (节点仓库 issue)[Bug] SSL: CERTIFICATE_VERIFY_FAILED when installing nodes in ComfyUI Manager (Windows Portable) ·
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/2076
- 166
- (官方 issue)We couldn't connect to 'https://huggingface.co' to load this file ... · Issue #5695 · Comfy-Org/Comf
- https://github.com/Comfy-Org/ComfyUI/issues/5695
- 167
- (官方 issue)PytorchStreamReader failed reading zip archive: failed finding central directory。How should this pro
- https://github.com/Comfy-Org/ComfyUI/issues/7185
- 168
- (中文社区)ComfyUI报错:huggingface_hub.local_entry_not_found - CSDN 问答
- https://ask.csdn.net/questions/8830209
- 169
- (官方文档)HfApi Client — RepositoryNotFoundError / RevisionNotFoundError / EntryNotFoundError / LocalEntryNotFound
- https://huggingface.co/docs/huggingface_hub/v0.35.0.rc1/package_reference/hf_api
- 170
- (节点仓库 issue)huggingface_hub.errors.RepositoryNotFoundError: 401 Client Error. · Issue #30 · lldacing/ComfyUI_P
- https://github.com/lldacing/ComfyUI_PuLID_Flux_ll/issues/30
- 171
- (节点仓库 issue)ComfyUI-Manager??why · Issue #1766 · Comfy-Org/ComfyUI-Manager
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/1766
- 172
- (节点仓库 issue)About Github Mirror · Issue #1697 · Comfy-Org/ComfyUI-Manager
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/1697
- 173
- (中文社区)【错误记录】从 GitHub 中 Clone 代码到本地报错 ( OpenSSL SSL_read: Connection was reset, errno 10054 )
- https://developer.aliyun.com/article/865207
- 174
- (中文社区)github拉取OpenSSL SSL_read: Connection was reset, errno 10054
- https://blog.csdn.net/weixin_47905944/article/details/150847977
- 175
- (节点仓库 issue)SOCKS proxy: `[WinError 64] The specified network name is no longer available` · Issue #1763
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/1763
- 176
- (官方文档)ComfyUI-Manager 配置(config.ini / network_mode / GITHUB_ENDPOINT / HF_ENDPOINT) - ComfyUI
- https://docs.comfy.org/zh/manager/configuration
- 177
- (社区)How to set startup proxy parameters for the desktop version of comfyui - ComfyUI Forum
- https://forum.comfy.org/t/how-to-set-startup-proxy-parameters-for-the-desktop-version-of-comfyui/742/6
- 178
- (中文社区)关于comfyui的pip设置镜像源以及torch缓存位置更改 - 知乎
- https://zhuanlan.zhihu.com/p/1946660222451052994
- 179
- (中文社区)pip install 太慢总是超时?国内镜像源配置的 4 种方法
- https://www.w3cschool.cn/article/63578772.html
- 180
- (官方 issue)listen argument does not work · Issue #3043 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/3043
- 181
- (官方 issue)unrecognized arguments: --listen · Issue #781 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/781
- 182
- (官方 issue)[Bug]: ComfyUI blank screen on loading after upgrading the frontend beyond 1.23.4
- https://github.com/Comfy-Org/ComfyUI_frontend/issues/4897
- 183
- (官方 issue)ComfyUI stuck loading forever since update
- https://github.com/Comfy-Org/ComfyUI/issues/7329
- 184
- (官方 issue)Freezing Interface & Workflow
- https://github.com/Comfy-Org/ComfyUI/issues/5805
- 185
- (官方 issue)Cannot execute because a node is missing the class_type property.: Node ID '#46'
- https://github.com/Comfy-Org/ComfyUI/issues/8903
- 186
- (官方 issue)Workflow does not load when nodes are missing, so the workflow cannot be fixed in the UI.
- https://github.com/Comfy-Org/ComfyUI_frontend/issues/6328
- 187
- (官方 issue)Loading Workflows with Missing Nodes
- https://github.com/Comfy-Org/ComfyUI/issues/6844
- 188
- (官方 issue)fronend has been buggiest than ever it doent even knows old workflows
- https://github.com/Comfy-Org/ComfyUI_frontend/issues/10506
- 189
- (官方 issue)[Feature Request]: Display warning when frontend version mismatches backend requirements
- https://github.com/Comfy-Org/ComfyUI_frontend/issues/4356
- 190
- (官方 issue)新版本前端与插件扩展不匹配的问题
- https://github.com/Comfy-Org/ComfyUI/issues/7258
- 191
- (社区)ComfyUI 工作流导入后,widget 消失了无法设置或者调整输入
- https://comfyui-wiki.com/zh/faq/widget-input-lost
- 192
- (官方 issue)TypeError: Failed to fetch
- https://github.com/Comfy-Org/ComfyUI/issues/9171
- 193
- (官方文档)Server Overview
- https://docs.comfy.org/development/comfyui-server/comms_overview
- 194
- (官方文档)Startup Flags
- https://docs.comfy.org/development/comfyui-server/startup-flags
- 195
- (官方 issue)Local API access
- https://github.com/Comfy-Org/ComfyUI/issues/11181
- 196
- (官方 issue)Load Image (From Output) has no Preview
- https://github.com/Comfy-Org/ComfyUI/issues/13078
- 197
- (官方 issue)Fix missing asset preview URLs outside input and output
- https://github.com/Comfy-Org/ComfyUI/pull/15509
- 198
- (社区)Setting storage for output bu v0.24.1 is ignored
- https://forum.comfy.org/t/setting-storage-for-output-bu-v0-24-1-is-ignored/4473
- 199
- (社区)ComfyUI 端口 8188 和 8000:局域网访问与远程访问设置
- https://comfyui-wiki.com/zh/faq/how-to-access-comfyui-on-lan
- 200
- (官方文档)ComfyUI Server Routes (HTTP and WebSocket API)
- https://docs.comfy.org/development/comfyui-server/comms_routes
- 201
- (官方 issue)Unable To Find Workflow in <filenamehere.png>
- https://github.com/Comfy-Org/ComfyUI/issues/7529
- 202
- (官方 issue)v0.3.49 will not load workflow from png if it was created with v0.3.48 and earlier
- https://github.com/Comfy-Org/ComfyUI/issues/9209
- 203
- (官方 issue)Since versions after 1.37.11 + the queue function, history, cancel button and "set the shapes" funct
- https://github.com/Comfy-Org/ComfyUI_frontend/issues/8911
- 204
- (官方 issue)[Unable to start] After installation, the startup remains in a spinning state.
- https://github.com/Comfy-Org/desktop/issues/587
- 205
- (官方 issue)Prompt outputs failed validation
- https://github.com/Comfy-Org/ComfyUI/issues/5403
- 206
- (官方文档)Comfy Settings
- https://docs.comfy.org/interface/settings/comfy
- 207
- (社区)Prompt outputs failed validation IPAdapterFluxLoader: - Required input is missing: ipadapter
- https://huggingface.co/InstantX/FLUX.1-dev-IP-Adapter/discussions/12
- 208
- (官方文档)ComfyUI LiteGraph (Canvas) Settings
- https://docs.comfy.org/interface/settings/lite-graph
- 209
- (官方文档)CUDA Toolkit Release Notes — Table 3/4 Corresponding Driver Versions
- https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
- 210
- (官方文档)CUDA GPU Compute Capability(GeForce RTX 5060 Ti 列在 Compute Capability 12.0 下)
- https://developer.nvidia.com/cuda-gpus
- 211
- (官方 issue)no kernel image is available for execution on the device · Issue #7677
- https://github.com/Comfy-Org/ComfyUI/issues/7677
- 212
- (官方 issue)Nvidia 50 Series (Blackwell) support thread: How to get ComfyUI running on your new 50 series GPU. ·
- https://github.com/Comfy-Org/ComfyUI/discussions/6643
- 213
- (官方 issue)CLIPTextEncode · Issue #11836(含同样 no kernel image 报错)
- https://github.com/Comfy-Org/ComfyUI/issues/11836
- 214
- (官方文档)Comfy-Org/docs PR #1491(新增该 Blackwell 条目)
- https://github.com/Comfy-Org/docs/pull/1491
- 215
- (官方文档)GPU Troubleshooting - ComfyUI(NVIDIA 各代卡的 SDPA / precision 排障)
- https://mintlify.wiki/Comfy-Org/ComfyUI/guides/gpu-troubleshooting
- 216
- (官方 issue)ComfyUI requires an NVIDIA GPU on Windows. No NVIDIA GPU was detected - ComfyUI 官方论坛(含 PATH 与 CIM 检查
- https://forum.comfy.org/t/comfyui-requires-an-nvidia-gpu-on-windows-no-nvidia-gpu-was-detected/686
- 217
- (官方文档)Command Line Arguments - ComfyUI
- https://mintlify.wiki/Comfy-Org/ComfyUI/configuration/command-line-args
- 218
- (官方 issue)Integrated GPU vs Discrete GPU · Issue #8848 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/8848
- 219
- (官方文档)Hardware Support - ComfyUI
- https://mintlify.wiki/Comfy-Org/ComfyUI/configuration/hardware-support
- 220
- (官方文档)ComfyUI README(AMD GPUs Windows/Linux 与 Intel GPUs 安装章节)
- https://raw.githubusercontent.com/comfyanonymous/ComfyUI/master/README.md
- 221
- (官方 issue)fix intel ARC "aten::_upsample_bicubic2d_aa.out" & Flux to cpu · Issue #4154
- https://github.com/Comfy-Org/ComfyUI/issues/4154
- 222
- (官方 issue)KSampler MPS backend out of memory · Issue #5810 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/issues/5810
- 223
- (社区)Node 'ID #267:275' has no class_type. The workflow may be corrupted or a custom node is missing.
- https://huggingface.co/Volkin/LTX-2.3-Workflows/discussions/1
- 224
- (官方 issue)Issue with workflow to mix human faces - Cannot execute because a node is missing the class_type pro
- https://github.com/Comfy-Org/ComfyUI/issues/11859
- 225
- (官方文档)Troubleshooting custom node issues - ComfyUI Docs
- https://docs.comfy.org/troubleshooting/custom-node-issues.md
- 226
- (节点仓库 issue)"Cannot connect to comfyregistry" - this is still throwing errors (despite being referred to as a
- https://github.com/Comfy-Org/ComfyUI-Manager/issues/2423
- 227
- (节点仓库 issue)IMPORT FAILED - ModuleNotFoundError: No module named 'ultralytics'
- https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues/793
- 228
- (节点仓库 issue)ComfyUI-Impact-Pack: (IMPORT FAILED) Subpack
- https://github.com/ltdrdata/ComfyUI-Impact-Pack/issues/689
- 229
- (社区)RunPod image: extra_model_paths.yaml omits ultralytics_bbox/ultralytics_segm/sams, so Impact Pack detector
- https://github.com/artokun/comfyui-mcp/issues/2302
- 230
- (节点仓库 issue)No module named 'timm'
- https://github.com/Fannovel16/comfyui_controlnet_aux/issues/20
- 231
- (节点仓库 issue)"Unsupported Nodes" after ComfyUI update
- https://github.com/Fannovel16/comfyui_controlnet_aux/issues/597
- 232
- (节点仓库 issue)comfyorg/comfyui-controlnet-aux
- https://github.com/comfyorg/comfyui-controlnet-aux
- 233
- (节点仓库 issue)WanVideoModelLoader - No module named 'sageattention'
- https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/10
- 234
- (官方 issue)WanVideoModelLoader: Can't import SageAttention after trying to install ultravico
- https://github.com/Comfy-Org/ComfyUI/issues/11195
- 235
- (节点仓库 issue)Flash Attention 2.8.3
- https://github.com/kijai/ComfyUI-WanVideoWrapper/issues/1317
- 236
- (节点仓库 issue)FFMPEG not found - ComfyUI Desktop
- https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/issues/349
- 237
- (节点仓库 issue)The only thing that comes to mind is that the system version of ffmpeg is installed is significant
- https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/issues/149
- 238
- (节点仓库 issue)h264 is not displayed in the format options of Video Combine.
- https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite/issues/188
- 239
- (中文社区)ComfyUI-VideoHelperSuite拓展报错ffmpeg is required for video outputs and could not be found.如何解决
- https://www.cnblogs.com/Coil1773647/p/18794634
- 240
- (节点仓库 issue)installing requirements says "Failed building wheel for insightface"
- https://github.com/hacksider/Deep-Live-Cam/issues/1578
- 241
- (社区)How to fix this issue "ERROR: Failed building wheel for insightface"
- https://stackoverflow.com/questions/76739044/how-to-fix-this-issue-error-failed-building-wheel-for-insightface
- 242
- (社区)Installation and Setup | Gourieff/ComfyUI-ReActor
- https://deepwiki.com/Gourieff/ComfyUI-ReActor/2-installation-and-setup
- 243
- (节点仓库 issue)Llama cpp quantize failed at GGML Assert
- https://github.com/city96/ComfyUI-GGUF/issues/327
- 244
- (节点仓库 issue)Wan 2.2 merge fails quantisation.
- https://github.com/city96/ComfyUI-GGUF/issues/351
- 245
- (节点仓库 issue)leejet/ComfyUI-GGUF: GGUF Quantization support for native ComfyUI models
- https://github.com/leejet/ComfyUI-GGUF
- 246
- (节点仓库 issue)IPAdapter model not found.
- https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/611
- 247
- (社区)How to fix: Error occurred when executing IPAdapterUnifiedLoaderFaceID: IPAdapter model not found?
- https://comfyui-guides.runcomfy.com/ultimate-comfyui-how-tos-a-runcomfy-guide/how-to-fix-error-occurred-when-executing-ipadapterunifiedloaderfaceid-ipadapter-model-not-found
- 248
- (中文社区)新版 ComfyUI IPAdapter plus 不再兼容旧版节点
- http://www.a-hxin.cn/archives/1141/
- 249
- (官方 issue)Error occurred when executing ADE_AnimateDiffLoaderWithContext:
- https://github.com/Comfy-Org/ComfyUI/issues/1840
- 250
- (节点仓库 issue)horrible results when using sd 1.5
- https://github.com/guoyww/AnimateDiff/issues/40
- 251
- (中文社区)AnimateDiff的motion module和base model分别在哪下载?
- https://ask.csdn.net/questions/9220249
- 252
- (社区)How do I fix "ComfyUI-Reactor node IMPORT FAILED" in ComfyUI?
- https://forum.smartart.live/topic/352-how-do-i-fix-comfyui-reactor-node-import-failed-in-comfyui/
- 253
- (社区)ComfyUI-Reactor README:inswapper_128.onnx 文件异常导致的报错说明
- https://huggingface.co/spaces/V0pr0S/ComfyUI-Reactor-Fast-Face-Swap-CPU/commit/35ba0ec134f63079a0cbbe67bcae92a2dacff1e8
- 254
- (官方仓库 PR)Make cu130 warning more visible. by comfyanonymous · Pull Request #15463 · Comfy-Org/ComfyUI
- https://github.com/Comfy-Org/ComfyUI/pull/15463
- 255
- (官方发布说明)Release v0.32.0 · Comfy-Org/ComfyUI(发布说明中含 Make cu130 warning more visible)
- https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.32.0
- 256
- (官方文档(本机实测 HTTP 200))CUDA Toolkit 13.0 Release Notes —— Minimum Required Driver Version for CUDA Minor Version
- https://docs.nvidia.com/cuda/archive/13.0.0/cuda-toolkit-release-notes/index.html
- 257
- (镜像站(本机实测 HTTP 200))阿里云 PyTorch wheel 镜像 cu130 目录(本机实测含 torch-2.11.0+cu130-cp312-cp312-win_amd64.whl)
- https://mirrors.aliyun.com/pytorch-wheels/cu130/
- 258
- (镜像站帮助页(本机实测))清华大学开源软件镜像站 —— PyPI 镜像使用帮助(本机实测 HTTP 200)
- https://mirrors.tuna.tsinghua.edu.cn/help/pypi/
- 259
- (代理服务(本机实测))ghproxy —— GitHub 文件代理(本机实测 HTTP 200,仅对 raw 文件有效)
- https://ghproxy.net/
- 附录 D 条目分类总览
- 类别
- 条目数
- 编号区间
- 安装、启动与运行环境
- 11
- 001 - 011
- 显存与内存不足(OOM)
- 11
- 012 - 022
- 依赖缺失与自定义节点安装失败
- 11
- 023 - 033
- 模型与权重文件加载
- 11
- 034 - 044
- 执行期运行时错误(张量/尺寸/节点)
- 11
- 045 - 055
- PyTorch、CUDA 运行库与加速库冲突
- 11
- 056 - 066
- 网络、下载与访问(含国内网络环境)
- 11
- 067 - 077
- 前端界面、浏览器与工作流加载
- 11
- 078 - 088
- 硬件、驱动、操作系统与非 N 卡环境
- 11
- 089 - 099
- 热门自定义节点/工作流专有报错
- 11
- 100 - 110
- 本机实测专项(RTX 5060 Ti / cu128 环境)
- 2
- 111 - 112
- 采集说明(各采集方向的覆盖范围与局限):
- • 安装、启动与运行环境:本次以 docs.comfy.org 官方文档、forum.comfy.org 官方论坛(含官方维护者 filtered / rhuang 的回复)、Comfy-Org/ComfyUI 官方仓库文件(run_nvidia_gpu.bat、pyproject.toml、README.md)为一手来源,辅以少量中文社区页面(fkaigc.com 报错库、ddd.cn 故障排除页、CSDN 问答标题)。局限:本次运行环境无法直连 github.com 网页与 GitHub API(被限流/超时),故 GitHub issue 只能引用搜索引擎返回的标题级信息,未取到 issue 正文原文;秋叶·绘世整合包未能取到可引用的报错原文,相关条目以中文社区标题级信息与通用机制描述呈现并已标注。
- • 显存与内存不足(OOM):本类目联网检索覆盖 ComfyUI 官方仓库 issue(Comfy-Org/ComfyUI,含 1918/5278/9121/9513/10475/10891/12713/13122/13954/14468/15443 等)、ComfyUI 官方排障与内存管理文档、ComfyUI 源码 comfy/cli_args.py 与 comfy/model_management.py(master 分支)、以及中文社区/云 GPU 文档页面。所有 sources 均为 web_search 或 GitHub API 实际返回的 URL。局限:GitHub API 匿名额度用尽(core 60/60),部分 issue 的评论正文未能取回,相关条目只依据 issue 正文与标题;NVIDIA/AMD 各型号的显存实测数值未逐一实机验证,凡未验证处均在 notes 标注 [未核实]。
- • 依赖缺失与自定义节点安装失败:本次覆盖 GitHub 官方/节点仓库 issue(ComfyUI、ComfyUI-Manager、kijai/WanVideoWrapper、cubiq/IPAdapter_plus、Fannovel16/controlnet_aux、Gourieff/ReActor、deepinsight/insightface、WanVideoWrapper、CogVideoXWrapper、artokun/comfyui-mcp、judian17/SDPose-OOD 等)、docs.comfy.org 官方排障与 Manager 文档,以及 20 余次中英文检索。局限:GitHub 网页域被本机网络阻断,issue 评论区正文未能全部抓取,部分条目的“解决步骤”来自同一 issue 的报告者/官方文档而非维护者回帖;中文社区来源(CSDN/gitcode 社区)可信度低于官方,仅作补充。 注:与第 10 类重复的「Manager 连不上 Comfy Registry」一条已在合稿时删除,保留第 10 类中更详细的版本。
- • 模型与权重文件加载:本类覆盖 safetensors 头部反序列化失败(HeaderTooLarge/MetadataIncompleteBuffer/incomplete metadata)、下载不完整与 Git LFS 指针、HTML 错误页伪装成权重、模型放错目录与 value not in list、LoRA 键不匹配、CLIP/CLIP Vision/VAE 缺失与错配、GGUF 与 fp8 量化加载、torch 2.6 weights_only 安全限制、Windows os error 1455 提交内存不足、文件名/中文路径问题。报错原文优先取自 ComfyUI 官方文档(zh/en)、forum.comfy.org、Hugging Face 讨论区与 node 仓库 issue;github.com 在本机网络下无法直连,因此部分 GitHub issue 仅作为搜索结果标题与链接引用,未能逐字抓取其正文。
- • 执行期运行时错误(张量/尺寸/节点):本次围绕 ComfyUI 执行期(点了运行之后才炸)的张量尺寸、数据类型、设备、节点返回空、采样器、NaN 出黑图等报错做联网采集,共约 20 次搜索,来源以 Comfy-Org/ComfyUI 与各节点仓库 issue、docs.comfy.org 官方排障文档、comfyui.nomadoor.net 社区笔记为主。局限:部分 issue 只有正文标题可核对(GitHub API 限流,未能逐条抓取全部评论),这类条目的 error_text 已标注 error_text_verbatim=false;个别 issue 的维护者最终结论未能核实,已在 notes 中标注 [未核实]。
- • PyTorch、CUDA 运行库与加速库冲突:本次采集以官方文档(docs.comfy.org、raw.githubusercontent.com 上的 ComfyUI/ComfyUI-Manager README)、torch/torchvision/xformers/SageAttention/triton-windows 仓库 README、PyPI 元数据、ComfyUI 官方论坛与若干社区排障文章为主;重点覆盖 RTX 50 系(Blackwell / sm_120)与 CUDA 12.8+ 时代。局限:本机网络无法访问 github.com 网页、Stack Overflow 与 Cloudflare 保护的站点,GitHub issue 只能通过搜索结果的标题/摘要片段与 api.github.com 逐条核实,故部分 issue 内的原始对话未能逐字取得;已尽量改用可完整抓取的官方文档与仓库 README 作为报错原文来源。所有版本号均来自实际抓取到的页面,未做推测。
- • 网络、下载与访问(含国内网络环境):本轮共联网检索 20 余次(英文+中文),来源以 github.com/Comfy-Org/ComfyUI、Comfy-Org/ComfyUI-Manager、Comfy-Org/desktop 的 issue、docs.comfy.org 官方文档、hf-mirror.com 官方说明页、RunPod 官方文档仓库、AutoDL 官方帮助文档、forum.comfy.org 为主,中文社区(CSDN/掘金/知乎/博客园)作为补充。所有 error_text 均摘自上述来源中真实出现的文本;无法逐字取到英文原文的条目已标注 "error_text_verbatim": false 并在 notes 中说明。局限:GitHub REST API 在检索中途触发限流(403),部分 issue 的解决回复未能逐条读取,相关条目的解决方案以官方文档与 issue 正文为准;Civitai 官方直链失效/401 的具体报错原文未取到逐字文本。
- • 前端界面、浏览器与工作流加载:本次覆盖 Comfy-Org/ComfyUI、Comfy-Org/ComfyUI_frontend、Comfy-Org/desktop 三个官方仓库的 issue/discussion、docs.comfy.org 官方文档(troubleshooting、custom-node-issues、comms_overview、comms_routes、startup-flags、settings)以及中文社区(CSDN 问答、ComfyUI Wiki、forum.comfy.org)。报错原文均取自 issue 正文/评论原文;命令行参数(--disable-all-custom-nodes、--front-end-version、--listen、--enable-cors-header、--output-directory)以 docs.comfy.org startup-flags 与 issue 评论原文为准。局限:GitHub API 在采集后段触发限流,ComfyUI#6451 / #5376 / #9635 / ComfyUI_frontend#8911 只拿到了 issue 标题与搜索结果摘要,未取到正文与维护者回复,相应条目已改用已取到原文的同类 issue 承载;官方文档目前没有为 /system_stats 与 /object_info 单独开页面,这两个接口的可用性依据是 issue 讨论中的实测引用(ComfyUI#7329 评论)与 ComfyUI 自身前端调用,未在官方 routes 文档页中列名。
- • 硬件、驱动、操作系统与非 N 卡环境:本次覆盖 NVIDIA 50 系(Blackwell sm_120)驱动/CUDA 门槛、no kernel image、cuDNN SDPA 失败、显卡不被识别/用了核显、多显卡选择、torch 未编译 CUDA、AMD(ROCm/DirectML)、Intel XPU、Apple MPS、CPU 模式、Windows 页面文件与长路径、模型放在其他盘、Docker/WSL2 GPU 透传、云 GPU 与 TDR 掉驱动。局限:GitHub REST API 本轮被限流,部分 issue 只能引用标题与 URL 而无法逐字抓取报错正文,涉及处已在 notes 标注 [未核实];docs.comfy.org 与 github.com 网页在本采集环境不可直连,官方文档正文取自其镜像站 mintlify.wiki 与 raw.githubusercontent.com,来源类型按官方文档标注。
- • 热门自定义节点/工作流专有报错:本次围绕 16 个最常被问到的热门节点包(Impact-Pack、Manager、AnimateDiff/Evolved、VideoHelperSuite、IPAdapter Plus、ControlNet Aux、insightface 系、SAM 与 Florence2、ComfyUI-GGUF、kijai/WanVideoWrapper、ReActor、WAS Node Suite、rgthree、Ultimate SD Upscale 等)做了 20+ 次中英文搜索,来源以各节点仓库与官方仓库 issue 为主。局限:web_search 返回的是标题+链接摘要而非 issue 正文,因此凡 error_text 取自 issue 标题的条目均在 notes 中标注来源形态,个别超长标题做了截断(截断处用…标出);未能取得正文原句的条目一律将 error_text_verbatim 标为 false。Windows 便携版 python_embeded\python.exe 路径在多个真实 install.bat / 自动安装脚本中出现,本节所有安装命令均以此为基准。
- • 本机实测专项(RTX 5060 Ti / cu128 环境):本类与其他类别不同:内容来自本机实测(启动日志、torch 查询、网络探测、镜像站 HTTP 校验),不是网络转述。每条都在 notes 里写明证据来源与未核实项,请按「本机」字样识别。
- 附录 E 数据来源与可信度说明
- 本手册共收录 112 条报错,引用 259 个来源页面。下面是关于这些内容从哪来、可信到什么程度的说明,请先读一遍再照着改环境。
- 一、来源类型分布
- 来源类型
- 页面数
- 说明
- GitHub(issue / PR / 仓库文件)
- 160
- 可信度最高:多为报错现场与维护者回复
- 官方文档 / 官方发布说明
- 48
- 可信度最高:行为定义与参数默认值以此为准
- 其他
- 24
- 需自行判断
- 中文社区文章 / 论坛
- 15
- 经验性解法,注意文章时效与版本差异
- HuggingFace(模型页 / 讨论区)
- 7
- 模型作者与用户讨论,可信度较高
- 英文社区论坛
- 3
- 经验性解法,需自行验证
- 开源镜像站
- 2
- 只用于说明下载通道,不代表技术结论
- 二、条目原文的完整度
- • 逐字报错原文:87 条。这些条目的「报错原文」是从 issue、官方文档或日志中照抄的真实文本,可以直接复制到搜索引擎或 Ctrl+F 检索。
- • 转述文本:25 条。采集时只拿到 issue 标题或社区转述,未能取得逐字原文,这些条目的原文块下方会标注「注:以上为转述文本,非逐字原文。」
- • 带 2 条以上来源的条目:108 条(共 112 条)。来源越多,说明该现象在不同环境被独立复现过,可信度更高。
- • 本机实测条目:2 条,单独归入「本机实测专项」章节,与网络来源严格分开标注,不在同一条里混用。
- 三、我们做了哪些验证,以及没能做什么
- • 已做(源码级核对):手册中关于启动参数语义的描述,已逐条与生成这台机器上的 C:\ComfyUI\ComfyUI\comfy\cli_args.py(ComfyUI 0.37.0 源码)核对,参数名与官方 help 原文一致。其中两条结论与网上大量旧教程相反,手册以源码为准:一是 --normalvram 在当前版本已不存在(写入启动脚本会报未识别参数),二是 dynamic vram 启用时 --lowvram 不产生任何作用。另外 --async-offload 的默认值 2,与本机启动日志 "Using async weight offloading with 2 streams" 相符。同理,ComfyUI-Manager 的 config.ini 键名(git_exe / security_level / network_mode)已与本机 Manager 源码 glob\manager_core.py 核对一致,config.ini 的两个版本路径差异也与本机启动日志打印的 config path 相符。
- • 已做(报错原文核对):对若干条标注为逐字原文的报错,在本机 ComfyUI 0.37.0 源码中核对了出处并逐字命中,例如 comfy\sd.py 中的 "Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding." 与 "ERROR: Could not detect model type of:"、comfy\model_management.py 中的 "Set vram state to:"、comfy\ldm\modules\attention.py 中的 "Using pytorch attention"、comfy\quant_ops.py 中的 "Failed to import comfy_kitchen, Error: ..., fp8 and fp4 support will not be available."。
- • 已做(抽样复核):对 GitHub 来源抽查了 5 条,用联网搜索核对 issue 编号与标题,4 条完全命中(标题与手册记录一致),1 条未能确认——该条所支撑的性能对比数字已在对应条目的「补充说明」里标注 [未核实]。未能确认不等于链接是假的:采集侧声称其 GitHub 链接均经 API 校验存在,而复核侧受本机网络限制无法直连 GitHub 复验,这一分歧无法在本机消除,故如实披露,请你按「待验证」对待该条数字。
- • 已做:本机环境数据(版本号、显卡、驱动、日志原文、网络可达性、镜像站与 wheel 文件的实际可下载性)全部来自本机实测命令与日志文件,未使用推断值。
- • 未做:生成本手册的机器直连 github.com、huggingface.co、docs.comfy.org 均超时(实测见「本机实测专项」章节),因此无法逐条实测每个来源链接当前是否可达。若你打开某个链接失败,先判断是不是网络原因,不要直接认为来源是假的。
- • 未做:所有解法都未在本机完整复现。手册里的步骤来自别人的现场,在你这台机器上可能需要微调路径、版本号与文件名。
- 四、使用这份手册的正确姿势
- • 改环境前先备份(尤其是 C:\ComfyUI\python 与 custom_nodes),升级 torch、批量重装依赖属于高风险操作。
- • 一次只改一个变量。同时做三件事再出问题,就无法判断是哪一步导致的。
- • 版本号要对照着看。同一条报错在 0.3.x 前后可能是完全不同的原因,每条都标注了「适用环境」与「版本」字段。
- • 拿不准就把完整 Traceback + 启动日志的版本信息一起贴出来求助,比只贴最后一行有效得多。
- 附录 F PyTorch / torchvision / torchaudio 配对对照表
- 「RuntimeError: Could not run 'torchvision::nms' with arguments from the 'CUDA' backend」「operator torchvision::nms does not exist」这类报错,绝大多数不是显存问题、也不是节点问题,而是这三个包版本没有配套。
- 下表不是从教程里抄的,而是从阿里云 PyTorch wheel 镜像实际下载对应 wheel 文件、读取其内部 METADATA 的 Requires-Dist 字段得到的(cp312 / Windows 构建),属于一手数据。
- 一、torchvision 与 torch 的硬绑定关系
- torchvision 版本
- 该 wheel 声明的依赖
- 说明
- torchvision 0.24.1
- torch == 2.9.1
- 对应 torch 2.9.x 线
- torchvision 0.26.0
- torch == 2.11.0
- ★ 本机当前组合,配对正确
- torchvision 0.27.1
- torch == 2.12.1
- 对应 torch 2.12.x 线(网上常见的「0.27.x 要配 torch 2.12+」说法与此一致)
- torchvision 0.28.0
- torch == 2.13.0
- 对应 torch 2.13.x 线
- 关键点:torchvision 的每个版本都精确绑定一个 torch 版本(== 而不是 >=)。所以「只升级 torch 不升级 torchvision」必然触发上面那两条报错;反过来,装 requirements.txt 时被顺带降级 torchvision,也会同样报错。
- 二、torchaudio
- torchaudio 的 wheel 元数据里没有声明对 torch 的依赖(实测其 Requires-Dist 不含 torch),所以它的配对靠版本号对齐——镜像目录里 torchaudio 的版本号与 torch 一一对应:2.9.1 / 2.10.0 / 2.11.0。按「与 torch 相同的版本号」来装即可。
- 三、三条铁律
- • 换 torch 时,torchvision 和 torchaudio 必须一起换,且版本号按上表对齐。
- • 用 --no-deps 安装,避免 pip 顺手改动其它依赖;装完立刻验证。
- • 别让任何自定义节点的 requirements.txt 决定 torch 版本——那是最常见的「装了个节点,整个 ComfyUI 崩了」的原因。
- 四、自己核实配对关系的方法(可复制执行)
- 不用相信任何教程,包括本手册:wheel 文件本身就是权威。用下面这段脚本读任一 wheel 的 METADATA 即可得到它的真实依赖。
- C:\ComfyUI\python\python.exe -c "import zipfile,sys;z=zipfile.ZipFile(sys.argv[1]);m=[n for n in z.namelist() if n.endswith('.dist-info/METADATA')][0];print('\n'.join(l for l in z.read(m).decode('utf-8','ignore').splitlines() if l.startswith('Requires-Dist') and 'torch' in l.lower()))" C:\path\to\torchvision-0.26.0+cu128-cp312-cp312-win_amd64.whl
- 已知的本机当前组合:torch 2.11.0+cu128、torchvision 0.26.0+cu128、torchaudio 2.11.0+cu128 —— 按上表核对,三者配对正确,所以本机目前的报错都不是版本不匹配引起的。
本条的可靠性边界
下面这些位置带
[未核实] 标记,是采集时没找到一手来源的部分,请自行验证:
- 补充说明
这条报错要用到的东西
按本条报错所属的类别(本机环境专项)列出站内相关的下载包与板块入口。 它们不是「必须下载才能修好」,而是这一类问题里最常要动到的东西;具体怎么做,以上面的解决步骤为准。
环境与依赖配置包依赖版本组合清单 / 环境自检脚本说明
待挂载
权重库核对自己手上权重文件的体积与目录
板块
同一类别的其他条目
这个解法对你有效吗
记录只存在你自己的浏览器里,不需要注册账号。
知仓学习社