ComfyUI-Nvidia-Docker mmartial/comfyui-nvidia-docker https://hub.docker.com/r/mmartial/comfyui-nvidia-docker ubuntu24_cuda12.6-latest Ubuntu 24.04 with CUDA 12.6 (current latest) -- last version to support GTX 10xx GPUs ubuntu24_cuda13.1-latest Ubuntu 24.04 with CUDA 13.1 ubuntu24_cuda13.0-latest Ubuntu 24.04 with CUDA 13.0 ubuntu24_cuda12.9-latest Ubuntu 24.04 with CUDA 12.9 ubuntu24_cuda12.8-latest Ubuntu 24.04 with CUDA 12.8 -- minimum CUDA version needed for Blackwell (inc RTX 50xx) hardware ubuntu24_cuda12.5-latest Ubuntu 24.04 with CUDA 12.5 ubuntu22_cuda12.4-latest Ubuntu 22.04 with CUDA 12.4 ubuntu22_cuda12.3-latest Ubuntu 22.04 with CUDA 12.3 ubuntu22_cuda12.2-latest Ubuntu 22.04 with CUDA 12.2 https://forums.unraid.net/topic/172874-support-comfyui-nvidia-docker/ https://github.com/mmartial/ComfyUI-Nvidia-Docker https://raw.githubusercontent.com/mmartial/unraid-templates/main/templates/ComfyUI-Nvidia-Docker.xml https://raw.githubusercontent.com/mmartial/unraid-templates/main/templates/img/ComfyUI-Nvidia-Docker.png AI: MediaApp:Photos WebUI images generation Stable Diffusion bridge **Nvidia Driver plugin** bash false ComfyUI WebUI Dockerfile with Nvidia support, installing ComfyUI from GitHub. Also installs ComfyUI Manager to simplify integration of additional custom nodes. The "run directory" contains HF, ComfyUI and venv. The "basedir" contains input, output and custom_nodes. The ComfyUI "Manager" is installed to simplify integration of additional custom nodes and their updates. The "Manager" is also used to update ComfyUI to more recent versions. All those folders will be created with the WANTED_UID and WANTED_GID parameters (by default using Unraid's default of 99:100) allowing the end-user to place directly into the folders their checkpoints, unet, lora and other required models. The container comes with no weights/models; you need to obtain those and install them in the proper directories under the mount you have selected for the "run directory". Output files will be placed into the basedir/output folder within the "basedir" directory. Please see https://github.com/mmartial/ComfyUI-Nvidia-Docker for further details. - Recommending setting USE_UV to true to use uv instead of pip - See details about "latest" tag - See details about "First time use" (and the "bottle" workflow), noting that Unraid's default YOUR_BASE_DIRECTORY should be /mnt/user/appdata/comfyui-nvidia/basedir - GTX 10xx users should use the ubuntu22_cuda12.6.3-latest tag and check the README.md for recommendations on how to run the container. - RTX 50xx (Blackwell) users should use at minimum ubuntu24_cuda12.8-latest tag. Note: - The container requires the Nvidia Driver plugin to be installed on your Unraid server. Usually that plugin will get you access to a CUDA driver with support for the latest tag available for this container. - This is a WebUI for the ComfyUI Stable Diffusion tool with a Docker image of usually over 4GB. - The container will take a while to start up, as it needs to download the ComfyUI Stable Diffusion tool and install its dependencies, usually adding another 5GB of downloaded content in the venv folder - The original Docker image is from Nvidia, as such it is governed by the NVIDIA Deep Learning Container License. - There are multiple version of the base image available, please select the one that fits your needs best. The name of the tag is the Ubuntu version followed by the CUDA version. Latest is set to point to the most recent combination as it should include the most recent software updates. For the complete list of supported versions, please see the GitHub repository https://github.com/mmartial/ComfyUI-Nvidia-Docker 2026-01-21 ### 20260121 - Adding TORCH_LOCK to lock a specific version of PyTorch, Torch Vision and Torch Audio -- Must set the CUDA backend to use. Must manually add the variable to use it on Unraid. ### 20260110 - Fixing git origin to point to the new repository ### 20260104 - Reversing git origin to point to the old repository until the transition is completed ### 20251230 - Fix pynvml uninstallation ### 20251229 - Changed git origin for ComfyUI - Renamed older images ### 20251228 - Address a FutureWarning: remove pynvml in favor of nvidia-ml-py - Note Ubuntu 22.04 images will be deprecated in 2026 (no date set yet) ### 20251219 - Added `USE_NEW_MANAGER` and `ENABLE_MANAGER_LEGACY_UI` to support new Manager introduced in Comfy 0.5. ### 20251211 - Updated ComfyUI Manager config.ini path: it is recommended to see Issue 92 for more details. ### 20251129 - Extended uv integration (including UV_TORCH_BACKEND) into userscripts_dir ### 20251122 - Added USE_UV environment variable to use uv instead of pip ### 20251005 - Fix loading configuration file overrides (e.g. comfyui-nvidia_config.sh) - Extended userscripts_dir to support Nunchaku and xformers ### 20251001 - Added additional libraries to the base image to support some often-used custom nodes - Added userscripts_dir script for onnxruntime ### 20250817 - Added automatic PyTorch selection for PREINSTALL_TORCH environment variable based on CUDA version - Added DISABLE_UPGRADES and PREINSTALL_TORCH_CMD environment variables ### 20250713 - Added PREINSTALL_TORCH environment variable which attempts to automatically select the cu128 index-url when CUDA 12.8 (or above) is detected. ### 20250607 - Added USE_PIPUPGRADE and USE_SOCAT environment variables - Added CUDA 12.2.2 and 12.9 images ### 20250503 - Future proof extras/PyTorch2.7-CUDA12.8.sh - Added run/pip_cache and run/tmp folders support ### 20250426 - Added `postvenv_script.bash` support - Blackwell users (RTX 50xx) must see README.md for details on how to enable PyTorch 2.7.0 with CUDA 12.8 support ### 20250424 - Removing CUDA 12.8/Blackwell special case: PyTorch 2.7.0 official is released - Only re-releasing CUDA 12.8 container - Added official 12.8 tag to Unraid release list ### 20250418 - use ENTRYPOINT to run the init script: allows for replacement of previously used command line arguments to support command line override - Added content in README.md to explain the use of comfytoo user - Added a section on reinstallation without losing our existing models folder. ### 20250413 - Make CUDA 12.6.3 the latest tag - Added support for /userscripts_dir and /comfyui-nvidia_config.sh (see README.md) ### 20250320 - Added CUDA 12.6.3 image, which will be the new latest as of the next release - Added checks for directory ownership - Added FORCE_CHOWN environment variable -- used to force change directory ownership as the comfy user during script startup (this process might be slow). This option was added to support users who mount the run and basedir folders onto other hosts which might not respect the UID/GID of the comfy user. - Added libEGL/Vulkan ICD loaders and libraries, including extension to the Windows usage section related to this addition ### 20250227 - Simplified user switching logic using the comfytoo user as the default entry point user that will set up the comfy user ### 20250216 - Fix issue with empty BASE_DIRECTORY ### 20250202 - Added BASE_DIRECTORY variable ### 20250116 - Added multiple builds for different base Ubuntu OS and CUDA combinations (with independent venv) - Added ffmpeg into the base container for some video workflows ### 20250109 - Integrated SECURITY_LEVELS within the docker arguments - Added libGL into the base container ### 20240915 - Added COMFY_CMDLINE_EXTRA variable to allow for additional command line arguments when starting ComfyUI (ex: --fast --reserve-vram 2.0 --lowvram). Check the Docker logs if this prevents your container from starting. ### 20240810 - Initial release http://[IP]:[PORT:8188] --runtime nvidia --gpus all 8188 /mnt/user/appdata/comfyui-nvidia/mnt /mnt/user/appdata/comfyui-nvidia/basedir 99 100 false true false /basedir normal true false true false personal_cloud