Skip to playerSkip to main content
Getting the dreaded "CUDA wasn't able to be loaded" error when installing ReActor Face Swapper on modern drivers? In this comprehensive troubleshooting guide, I show you exactly how to install the ReActor custom node inside ComfyUI Portable when running modern CUDA 12.x environments (instead of the older CUDA 11.8 standard).

I walk you step-by-step through checking your embedded Python version, manually downloading and installing the matching pre-compiled library wheels (.whl) for your portable directory, and upgrading PIP. Then, I tackle the core environment conflict: manually installing NVIDIA cuDNN 9, copying the essential binaries into your CUDA Toolkit paths, and forcing the installation of the specific ONNX Runtime GPU package configured for CUDA 12. Finally, we build a clean test workflow connecting ReActor Fast Face Swap with your input and source image paths to verify a successful face swap.

Original YouTube Tutorial: https://youtu.be/zCvik-ohKko

Connected Videos in this Series:
* How to Install ComfyUI: https://youtu.be/ynfNJEtvUtQ
* How to Install Git: https://youtu.be/ae1HrqDRv4E
* Fix Missing Nodes with ComfyUI Manager: https://youtu.be/dyrhPVRsy9w
* ComfyUI Video Face-Swapping and Upscaling Guide: https://youtu.be/aJwl7z-1iB8
* learn more about face restore GFPGAN and Codeformer: https://youtu.be/rTIxH4pJv4g

Face Swaper Alternatif using Automatic1111 or Forge
Roop: https://youtu.be/9cnWU0HVbuY
Control Net Face Swaper: https://youtu.be/aJwl7z-1iB8

Source/Model Resources:
* ComfyUI ReActor Node GitHub Repository: https://github.com/Gourieff/comfyui-reactor-node
* NVIDIA cuDNN Archive Downloads: https://developer.nvidia.com/cudnn
* Official ComfyUI Hub: https://github.com/comfyanonymous/ComfyUI

Important script:
git clone script in custom nodes: git clone https://github.com/Gourieff/comfyui-reactor-node.git
uninstall onyx
pip uninstall onnxruntime-gpu

install onyx to use cuda 12.x version script
pip install onnxruntime-gpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/

Video Details:
* Original Publish Date: July 10, 2024
* Focus: ReActor Face Swapper / CUDA 12 Environment Fixes / ONNX Runtime GPU Troubleshooting
* Testing Environment: ComfyUI Portable, Windows 11, NVIDIA RTX GPU

Follow lordcaocao2025 on Dailymotion for more advanced technical troubleshooting, custom node configurations, and seamless local AI pipeline setup guides!

---
Connect with me:
๐Ÿ“บ YouTube: https://www.youtube.com/@CaoCao2025
๐Ÿ“ฑ TikTok: https://www.tiktok.com/@caocao20250
๐Ÿ’Ž Patreon: https://www.patreon.com/cw/Caocao2025

#ComfyUI #ReActorFaceSwap #CUDA12 #ONNXRuntime #ComfyUIPortable #FaceSwapper #AITutorial #lordcaocao2025
Transcript
00:00Hello guys, welcome back with me, ChowChow2025.
00:04Today we're going to install Reactor Face Swapper for ComfyUI.
00:08What we need first is Git, ComfyEye and CUDA.
00:12How to install them could refer on the link on the description.
00:16In this video I use CUDA 12.1, so we're going to face a lot of problem in this installation,
00:24because this thing was supposed for CUDA 11.8.
00:27First, access the GitHub page, link on the description.
00:32Next, click the code button and copy the script or link.
00:37Access the ComfyUI custom node directory and open command prompt from there.
00:42The directory is comfyui slash custom underscore nodes.
00:48From the command prompt, write git clone and paste the link with right click.
00:52You could also copy-paste it from the description.
00:55After it finished, a new folder called ComfyEye Reactor Node will be created.
01:01Access that folder.
01:04You will find install.bat there.
01:07Double-click it to execute installation process.
01:14As you can see, we've finished installed Onyx 1.16.
01:18The problem is 1.16 is not for our CUDA.
01:22Also, if you have problem other than that, you could install wheel manually by downloading wheel for your portable Python.
01:29Check the portable version Python by opening command prompt from ComfyPortableRoute directory.
01:35Write python-in-beaded slash python dot exe space dash v.
01:41Finally, based on your Python version, copy one of the link in the GitHub.
01:46Download them at the ComfyUI PortableRoute directory.
02:02Update your pip using the script using python-in-beaded slash python dot e xa dash e dash mpipe install
02:10dash upipe.
02:12And do pip install them by copying the script on the GitHub page.
02:27After that, go back to ComfyEye slash custom underscore nords and rerun the install dot bat.
02:33I don't think this is necessary, but I do it anyway.
02:41Now we could try to run ComfyEye.
02:44It will download and set up a bunch of things.
02:47If you're ComfyEye startup without node, don't worry.
02:51It's simply just downloading file and the system not fully set up yet.
02:55A few minutes will pass and finally nodes will be showing up.
02:59Clear all the nodes.
03:01Add nodes from reactor, reactor, fast face, swap.
03:11Add two load image nodes.
03:21Also add preview image for the result.
03:29Connect the image to input image and source image.
03:32Connect the reactor image result with preview image.
03:36Input image is the image that the face going to be changed.
03:40Source image is the face that will be put on the input image.
03:44Refer face restore model to my upscaling video.
03:48Link on the description.
03:50Basically, it's just to decide how well the face will be fixed.
03:54The more the visibility and strength, the better the face will be,
03:57but it will also reduce the similarity of the source image face.
04:02Run it by clicking Q prompt.
04:06It will first start downloading model.
04:09In my case, since we used the wrong CUDA, we will met this error.
04:14CUDA path is set, but CUDA wasn't able to be loaded.
04:17Please install the correct version of CUDA and CUDNN as mentioned in the GPU requirements page.
04:24A link in bracket.
04:25Make sure they're in the path and that your GPU is supported.
04:28I'm not sure which step able to solve this, but here what I do.
04:34First, I install CUDNN from NVIDIA site.
04:38Link on the description.
04:39Refer to support matrix between CUDA and CUDNN version.
04:43In this case, since I use CUDA 12, next I download CUDNN version 9.
04:50Download the CUDNN library and install it.
05:06If you are asked to restart your computer first, do so.
05:11After that, just install it like regular installation process.
05:14Now access your CUDNN installation directory.
05:19Before you do this, I suggest you make a backup of you, CUDA Toolkit, first.
05:24Copy all the folder in the bin, include and lib directory for 12.3 to your CUDA Toolkit folder.
05:32Paste it in the same bin to bin, include to include and lib to live.
05:38Now access your CUDNN installation directory.
06:18The next step is to uninstall Onyx Runtime and force it to install Onyx Runtime GPAR
06:24for CUDA 12.
06:25Copy the script from the video description.
06:32After some long loading.
06:35Now let's try again.
06:42Quill, prompt, and here are the results.
07:27Here, another example.
07:45Anyway, guys, that's it for today video.
07:47Thank you for watching Don't forget to like, share, and comment.
07:54See you again on the next episode.
07:57And have a nice day.
08:00Please subscribe.
Comments
lordcaocao2025
Creator
Getting ReActor to play nice with CUDA 12.x in ComfyUI Portable is a headache since the default script targets 11.8. Forcing the correct ONNX Runtime GPU wheel and dropping the cuDNN 9 binaries directly into the toolkit paths completely patches the startup crash. Did your node finally load without throwing the CUDA error? Drop a comment below if you hit any roadblocks!

Recommended