Just like ComfyUI official Standalone Portable, but preloaded with numerous custom nodes and Python packages, with all dependencies resolved.
The rest of README was translated by ChatGPT and edited by me (how lazy, how good😙).
-
Pre-installed with 30+ custom nodes including:
-
ComfyUI-3D-Pack which you may want to spend less time on installing.
-
Nodes like ReActor Node that download model files on the first run (often causing freezing for users with bad Internet Connection)
-
-
Includes 240+ Python packages, versions compatible, including:
-
Packages like
insightface
anddlib
which often throw "compiler not found" during installation -
Packages like CLIP that require compilation from a git repository
-
ONNX Runtime, which often throws "CUDAExecutionProvider Not Available". The one in the package not only works, but also supports CUDA 12.x
-
xFormers. I know it’s not included in the ComfyUI official package for a good reason. See details in About xFormers.
-
-
Only a few functional models that might slow down the program startup are pre-installed; users please prepare SD models of your favorite.
-
Only NVIDIA GPUs are supported, preferably with the latest drivers installed.
-
AMD GPUs may possibly work with ZLUDA after installing the HIP SDK, and this requires modifying startup parameters, among other things.
-
-
Download the package files from the release page:
-
https://github.com/YanWenKun/ComfyUI-Windows-Portable/releases
-
Note that there are multiple files (compressed in volumes) that need to be downloaded and extracted.
-
-
Place SD models in
ComfyUI\models\checkpoints
.-
Subdirectories can be created for better organization.
-
-
Run
run_nvidia_gpu.bat
to start ComfyUI. -
After starting, the program will automatically open a browser, or you can visit: http://localhost:8188/
-
If you don’t want the browser to open automatically, edit
run_nvidia_gpu.bat
and add the startup parameter--disable-auto-launch
.
-
Few steps before running Comfy3D:
-
Install Visual Studio Build Tools 2022 with "Desktop development with C++" workload. Its compiler will be needed (for some JIT code).
-
Install CUDA Toolkit, latest version is fine as long as it’s 12.x.
-
Edit
run_nvidia_gpu.bat
, reduceTORCH_CUDA_ARCH_LIST
according to your GPU arch, to save your time on JIT compiling.
-
Model Path: If model files are already saved in another location and you don’t want to move them, you can configure the program to load them:
-
Rename
extra_model_paths.yaml.example
in theComfyUI
directory toextra_model_paths.yaml
. -
Edit
extra_model_paths.yaml
according to the directory structure, removing corresponding comments. -
Items other than
base_path
can be added or removed freely to map newly added subdirectories; the program will try load all of them.
-
-
Proxy: If a network proxy is needed, edit
run_nvidia_gpu.bat
and add the following at the beginning:set HTTP_PROXY=http://localhost:1081 set HTTPS_PROXY=http://localhost:1081
-
Manager: Take advantage of the ComfyUI-Manager (the Manager button located at the bottom right corner of the ComfyUI page) to install additional custom nodes, update ComfyUI and nodes, or disable unnecessary nodes.
-
If you find that some unused nodes affect the startup time, you can disable them.
-
-
Language: Click the gear (⚙) icon at the top right corner of the ComfyUI page to modify settings. Find AGLTranslation to change the language (default is English, options are {Chinese, Japanese, Korean}).
-
How to upgrade: ComfyUI-Manager can do most updates, but if you want a "fresh" upgrade, you can first delete the
python_embeded
directory, and then extract the same-named directory from the new version’s package to the original location. Additional "try fix" in ComfyUI-Manager may be needed.
Most of the popular custom nodes are denpendency-compatible. You can install more custom nodes through ComfyUI-Manager.
If compatibility issues arise, you can try disabling conflicting nodes in ComfyUI-Manager.
The cross-attention mechanism in PyTorch 2.2+ performs well on Windows and is more consistent (albeit subtly), so it’s understandable why the ComfyUI official package chose not to include xFormers.
Speed and memory usage of xFormers and Torch vary depending on the application scenario and require specific testing. You can add --use-pytorch-cross-attention
to the ComfyUI startup parameters to disable xFormers.
Personally, I prefer to enable xFormers when generating videos.
Additionally, including xFormers satisfies the dependencies of some nodes like AnimateAnyone Evolved.
This repository utilizes a pipeline to build the integration package, and the codebase doesn’t contain specific configurations or require additional access permissions. Hence, you can directly fork this repository to start executing the GitHub Workflow.
-
After forking, go to Actions on the page.
-
Locate Build & Upload Package.
-
For example, it looks like this on my repository page.
-
-
Click Run Workflow.
-
Wait about 20~40 minutes until the workflow run complete.
-
Go to the releases page of your repository, where you’ll find the newly generated draft for download or publish.
Files detected by Sandboxie as shown in the image; registry changes are not clear.
If you need to configure a sandbox, it is recommended to set the program directory (the parent directory of ComfyUI) to "Full Access" under "Resource Access."
In my personal experience, I use a sandbox not so much for security considerations but mainly to avoid various Python packages downloading files haphazardly. Especially Huggingface Hub likes to download to %USERPROFILE%\.cache
, and some rough and ready nodes directly call its downloader, downloading Git LFS blob files rather than individual model files, which are neither intuitive nor convenient for copying and reusing. Of course, despite the complaints, sandboxing is mainly for easy cleaning of temporary files.
Linux/WSL2 users may want to check out my ComfyUI-Docker, which is the exact opposite of the Windows integration package in terms of being large and comprehensive but difficult to update. I designed the Docker image with a meticulous eye, selecting a series of non-conflicting and latest version dependencies, and adhering to the KISS principle by only including ComfyUI-Manager, leaving the choice of custom nodes to the users. Not to mention the benefits of easy upgrades, easy cleanup, and secure isolation provided by container runtime itself.
Originally, the code was copied from ComfyUI’s GitHub workflow, but I found it difficult to debug, so I rewrote the script.
However, the packaging concept is similar: providing a self-contained, portable, and fully equipped environment with a pre-installed Python Embedded.
The difference is that I didn’t download wheels first and then install them in bulk like comfyanonymous did. Because dependency relationships are too tricky, I went straight to pip install
for dependency solving.
Thanks to the ComfyUI GitHub workflow, from which I drew inspiration. The initial code was also copied from there.