ComfyUI for Linux (Install Script) / Windows (Portable Version)

Linux

• Readme (github)

• first open the script in a text editor and adjust the pathes and versions, currently its:

  • PYTHON_VERSION="3.12.10" # Python version to install via pyenv

  • VENV_PATH="/mnt/daten/AI/comfy_env" # Virtual environment location

  • PYTORCH_VERSION="2.9" # PyTorch major.minor version for wheel URLs (e.g., 2.9, 2.10)

  • PYTORCH_FULL_VERSION="2.9.1+cu128" # Full PyTorch version for pip install e.g., 2.9.1+cu128, 2.10.0+cpu)

  • PYTORCH_INDEX_URL="https://download.pytorch.org/whl/cu128" # PyTorch index URL (cu128, cu121, cpu)

  • NUMPY_VERSION="2.2.6" # NumPy version (2.2.x compatible with PyTorch 2.9+)

  • TRANSFORMERS_VERSION="4.57.3" # Transformers version (5.x for Mistral3/Ministral3 support, breaks florence2)

  • COMFYUI_FRONTEND_VERSION="1.36.13" # ComfyUI frontend version (installed in venv's site-packages)

  • # ComfyUI installation configuration

  • COMFYUI_PARENT_DIR="/mnt/daten/AI" # Parent directory where ComfyUI will be cloned

  • COMFYUI_DIR_NAME="ComfyUI" # ComfyUI folder name (change if you want a different name or multible versions)

  • COMFYUI_VERSION="v0.8.2" # ComfyUI version to checkout (tag, branch, or commit SHA). Leave empty for latest.

  • CREATE_SYMLINKS=true # Set to false to skip symlink creation

  • USER_MODELS_PATH="/mnt/daten/AI/models" # Your centralized models directory

  • USER_OUTPUT_PATH="/mnt/daten/AI/output" # Your centralized output directory

  • USER_CUSTOM_NODES_PATH="/mnt/daten/AI/custom_nodes" # Your centralized custom_nodes directory (shared across ComfyUI installations)

  • INSTALL_NUNCHAKU=true # Set to false to skip Nunchaku (NVIDIA GPU required)

• to execute the script open a bash in the folder where you've extracted the file and type chmod +x install_comfy_env.sh after that you can execute it with a right click of your mouse on the script and select execute as app.

• if you do all steps (default) the script will:

  • install pyenv (if not existing),

  • the selected python version,

  • create the comfy_env (venv) and

  • installs comfyui.

  • it also clones a set of custom nodes that are needed for my workflows (check the script, the block starts at line 551).

• in its current state it installs python 3.12.10, pytorch 2.9.1 cu128, nunchaku 1.2.1 package, facexlib, insightface etc. it installs the requirements of all cloned custom nodes.

• it also adds code to the .bashrc if selected. that way you can start comfyui faster by typing comfyui in the bash. but it also creates a startup script that allows you to start comfyui (right click on start-comfyui.sh and execute as app).

• the final steps in the script is forcing the versions that are setup at the beginning. e.g. pytorch 2.9.1, transformers 4.57.3, numpy 2.2.6 because those are most likely overwritten by the requirements of the custom nodes (like reactor, facenet_pytorch etc.).

• for the different distros there is a package manager detection so it should work no matter what distro you use. this is only used for the pyenv/python setup in Step 1 all other installations are done with pip.

  • Debian/Ubuntu: apt-get

  • Fedora/RHEL: dnf (my system)

  • CentOS/older RHEL: yum

  • Arch Linux: pacman

  • openSUSE: zypper

• in the last update i've added step 11 which creates 2 aliases now

  • comfyui (starts comfyui directly)

  • envact (activates the venv)

• I've also added a Multi-shell detection to write the aliases to the correct files:

  • Bash → adds aliases to ~/.bashrc

  • Zsh → adds aliases to ~/.zshrc

  • Fish → adds functions to ~/.config/fish/config.fish (Fish uses different syntax)

  • Fallback → uses ~/.profile if no common shells detected

• since im not using windows anymore there will be no further updates for the portable versions 🤷🏻.

Version 0.8.2

• pyTorch 2.9+cu128

• python: 3.12.10

• Frontend: 1.36.13

• removed crystools -> replaced with crystools-monitor only

• Even though some API imports have been fixed, the deprecated messages are still displayed.

• the reason for adding the forked versions instead of the originals is that some of the nodes are currently not working with the latest comfyui version.

Additional Installed Libs

• triton 3.51.1

• llama_cpp_python 0.3.16

• flash attention 2 (v2.8.3+torch2.9+cu128-cp312)

• nunchaku-1.0.2+torch2.9-cp312-cp312

• sage attention 2.2.0

• transformers==4.56.2

• insightface, facexlib, facenet_pytorch, comfy-cli, bitsandbytes, tensorflow, hf_xet, requests, pilgram, tf-keras, setuptools_scm, qwen-vl-utils, huggingface-hub, accelerate, Pillow, opencv-python, safetensors, psutil, PyYAML, aiohttp, av, gguf, ultralytics, pytorch-extension

• python headers for torch compile/triton

Installed Custom Nodes

• comfy-mtb

• comfyui_birefnet_ll

• comfyui_controlnet_aux

• ComfyUI_Eclipse 1.2.1

• comfyui_essentials

• comfyui_essentials_mb

• comfyui_layerstyle

• ComfyUI_LayerStyle_Advance

• ComfyUI_Patches_ll (forked and fixed Imports)

• comfyui_pulid_flux_ll

• comfyui_ttp_toolset

• comfyui-advanced-controlnet

• ComfyUI-Crystools-MonitorOnly (forked and fixed small error in browser log)

• ComfyUI-Custom-Scripts (forked and fixed API Imports)

• comfyui-detail-daemon

• comfyui-easy-use

• comfyui-florence2

• comfyui-frame-interpolation

• ComfyUI-GGUF

• ComfyUI-GIMM-VFI

• comfyui-impact-pack

• comfyui-kjnodes

• ComfyUI-Manager (removed the anoying fetch msg output)

• ComfyUI-nunchaku (forked and fixed Imports)

• ComfyUI-ReActor

• comfyui-supir

• ComfyUI-TeaCache (forked and fixed Imports)

• comfyui-videohelpersuite

• ComfyUI-WanVideoWrapper

• raffle

• RES4LYF (forked and partially fixed (community pull requests))

• rgthree-comfy (forked and fixed API Imports)

• sd-dynamic-thresholding

• was-ns

Version 0.3.75

(You will see some warnings from various nodes because the authors of comfyui are changing the API. -> [DEPRECATION WARNING])

• Frontend Version: 1.32.3 (fix)

  • i've uploaded a patch for the latest frontend version if you want to upgrade to 1.34.3:
    https://civarchive.com/articles/23392/patch-for-the-comfyui-frontend

• pyTorch 2.9+cu128

• python: 3.12.10

Additional Installed Libs

• triton 3.51.1

• llama_cpp_python 0.3.16

• flash attention 2 (v2.8.3+torch2.9+cu128-cp312)

• nunchaku-1.0.2+torch2.9-cp312-cp312

• sage attention 2.2.0

• transformers==4.56.2

• facenet_pytorch, pytorch-extension, facexlib, insightface, comfy-cli, bitsandbytes, tensorflow, hf_xet, requests, pilgram, tf-keras, nvidia-ml-py

• python headers for torch compile/triton

Installed Custom Nodes

• comfy-mtb

• comfyui_birefnet_ll

• comfyui_controlnet_aux

• ComfyUI_Eclipse (1.0.60)

• comfyui_essentials

• ComfyUI_essentials_mb

• comfyui_layerstyle

• ComfyUI_LayerStyle_Advance

• comfyui_patches_ll

• comfyui_pulid_flux_ll

• comfyui_ttp_toolset

• comfyui-advanced-controlnet

• ComfyUI-Crystools

• comfyui-custom-scripts

• comfyui-depthanythingv2

• comfyui-detail-daemon

• comfyui-easy-use

• comfyui-florence2

• comfyui-frame-interpolation

• ComfyUI-GGUF

• ComfyUI-GIMM-VFI

• comfyui-impact-pack

• comfyui-impact-subpack

• comfyui-inspire-pack

• comfyui-jdcn

• comfyui-kjnodes

• ComfyUI-Manager

• comfyui-multigpu

• ComfyUI-nunchaku

• comfyui-reactor

• comfyui-supir

• comfyui-videohelpersuite

• ComfyUI-WanMoeKSampler

• ComfyUI-WanVideoWrapper

• facerestore_cf

• raffle

• RES4LYF

• rgthree-comfy

• sd-dynamic-thresholding

• sdxl_prompt_styler

• teacache

• was-ns

• whiterabbit