现在的趋势是AI模型越来越强,生图、生视频从SD、MJ到即梦、香蕉、豆包等等,五花八门,普通人使用AI的门槛越来越低,现在只需一句提示词即可生成想要的图片和视频,很多人误判ComfyUI要被淘汰了,认为节点工作流没用了,但我认为专业生产端反而越来越依赖ComfyUI,所以问题不在于:ComfyUI会不会死,而在于:你属于哪一类用户,一句话出图≠生产系统!
ComfyUI本质上不是一个“绘图软件”,像Midjourney,他更像一个推理编排系统,更适合:电商商品工厂、企业级AI内容生产、批量自动化、本地私有化部署、多模型融合、深度统一控制、商业化。文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
DGX Spark本地化部署:文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
仅适用于文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
*NVIDIA DGX Spark文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
*DGX OS (Ubuntu ARM64)文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
* 本地AI生图/电商图/Flux/SDXL/Wan视频工作流文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
部署步骤:
环境部署
创建目录和权限
sudo mkdir -p /opt/ai/{models,workflows,outputs,hf-cache,venvs}
sudo chown -R $USER:$USER /opt/ai
验证环境
python3 –version nvidia-smi nvcc –version
如果提示cuda坏境或组件不存在就重新安装文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
sudo apt install nvidia-cuda-toolkit -y
创建ComfyUI虚拟环境
创建虚拟环境
cd /opt/ai/venvs python3 -m venv comfyui
激活虚拟环境
source /opt/ai/venvs/comfyui/bin/activate
升级虚拟环境
pip install -U pip setuptools wheel
安装 DGX Spark 专用 PyTorch(核心)
安装PyTorch
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu130
验证
python -c "import torch;print(torch.cuda.is_available())"
正常应输出:True文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
python -c "import torch;print(torch.cuda.get_device_name(0))"
正常应输出:NVIDIA GB10文章源自八点运动-https://www.8oio.com/comfyui-for-dgx-spark-%e4%bc%81%e4%b8%9a%e7%94%9f%e4%ba%a7%e7%ba%a7%e4%bf%9d%e5%a7%86%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b/530/
克隆ComfyUI
克隆ComfyUi
cd /opt/ai git clone https://github.com/comfyanonymous/ComfyUI.git
或国内网络
git clone https://gitee.com/mirrors/comfyui.git
安装依赖
cd comfyui pip install -r requirements.txt
安装Triton
pip install triton
安装加速组件
pip install \ accelerate \ safetensors \ sentencepiece \ einops \ transformers \ diffusers \ opencv-python
安装 ComfyUI-Manager(强烈建议)
它用于:
* 自动安装节点
* 自动依赖管理
* 工作流兼容
* 模型下载
这是生产环境必装。
cd /opt/ai/comfyui/custom_nodes git clone https://github.com/ltdrdata/ComfyUI-Manager.git
或国内网络
git clone https://ghfast.top/https://github.com/ltdrdata/ComfyUI-Manager.git git clone https://mirror.ghproxy.com/https://github.com/ltdrdata/ComfyUI-Manager.git
配置模型目录(推荐独立)
生成模型目录
mkdir -p /opt/ai/models/{checkpoints,configs,vae,loras,upscale_models,embeddings,hypernetworks,controlnet,clip,clip_vision,style_models,diffusion_models,unet,vae_approx}
编辑生成配置文件
cd /opt/ai/comfyui nano extra_model_paths.yaml
写入
spark_models: base_path: /opt/ai/models checkpoints: checkpoints configs: configs vae: vae loras: loras upscale_models: upscale_models embeddings: embeddings hypernetworks: hypernetworks controlnet: controlnet clip: clip clip_vision: clip_vision style_models: style_models diffusion_models: diffusion_models unet: unet vae_approx: vae_approx
Ctrl+O & Ctrl+X 保存退出
DGX Spark 最佳启动参数(核心)
创建启动脚本
nano /opt/ai/start-comfy.sh
写入
source /opt/ai/venvs/comfyui/bin/activate cd /opt/ai/comfyui python main.py \ --listen 0.0.0.0 \ --port 8188 \ --fast \ --use-pytorch-cross-attention
授权启动脚本
chmod +x /opt/ai/start-comfy.sh
执行启动
/opt/ai/start-comfy.sh
自动启动(生产级)
实现DGX Spark开机后自动启动服务
创建 service:
sudo nano /etc/systemd/system/comfyui.service
写入
[Unit] Description=ComfyUI Service After=network.target [Service] Type=simple User=mruei WorkingDirectory=/opt/ai/comfyui ExecStart=/bin/bash /opt/ai/start-comfy.sh Restart=always RestartSec=10 [Install] WantedBy=multi-user.target
保存退出
检查启动脚本是否存在:
ls -l /opt/ai/start-comfy.sh
必须有X权限,没有就执行:
chmod +x /opt/ai/start-comfy.sh
启用自动服务
sudo systemctl daemon-reload sudo systemctl enable comfyui sudo systemctl start comfyui
查看是否运行
systemctl status comfyui
创建运行日志(运行更加稳定)
创建日志目录
mkdir -p /opt/ai/logs
修改启动脚本并增加日志功能
nano /opt/ai/start-comfy.sh
写入
source /opt/ai/venvs/comfyui/bin/activate export HF_HOME=/opt/ai/hf-cache export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True cd /opt/ai/comfyui python main.py \ --listen 0.0.0.0 \ --port 8188 \ --fast \ --use-pytorch-cross-attention \ >> /opt/ai/logs/comfyui.log 2>&1
日常维护
日常更新
cd /opt/ai/comfyui git pull
日常监控
htop
GPU
watch -n 1 nvidia-smi










