Better ComfyUI Slim on RunPod - August 2025 Version + Manager + Crystools + Instructions for Newbies
Container image link docker.io/unaico/comfypod:latest
This is a fork of kodxana's original Dockerfiles for a slim ComfyUI build. I was having various issues getting it to run in August 2025, so this version is hardcoded with a working dependency list as of 8/5/2025, and lots of debugging info in the console logs if you run into any issues. The Crystools installation still doesn't seem correct - you will want to go to the Comfy Manager and reinstall Crystools.
The original version had a 5090 version - I do not. This is for non-5090 CUDA devices like A100s, H100s, H200s, B200s, etc etc. The original version also had a File Browser and Zasper, those have also been left out.
After ComfyUI boots up, RunPod still needs a few minutes until you can access it via the web. If you see the message "Using pytorch attention" - just wait about 2 minutes. The final message you will get from the Container Log confirming it's ready is this: "To see the GUI go to: http://0.0.0.0:8188"
I had a lot of questions about downloading and storing models. Originally I was trying to use Cloudflare R2 as storage, but ultimately I could not figure it out. Doing the storage via RunPod makes things a lot easier. Start by making a Network Volume with about 150 GB. You will need to know which sort of GPU you are planning on using because each storage location has certain machines available to it with the storage. In my case I wanted to use A100s, so I chose the Kansas location as that's where the most machines were.
From now on, any time you start anything on RunPod, start by going to your Storage, clicking your new Network Volume, and then clicking Deploy Pod With Volume. The first time you connect with the network volume, it will kick off an install script that might take 30 minutes or so. This is installing ComfyUI to your network volume.
As far as downloading the models/safetensors/checkpoints - cURLing via the Web SSH is by far the easiest and most effective method, and 5x faster than runpodctl. The easiest way to find out what you need to download is to just launch ComfyUI, click File -> Browse Templates, and choose whatever you'd like. These days, their template gallery is immediately updated with new models. So for example if you wanted to use Flux Dev you would click that one, and then ComfyUI will tell you which files you need, what folder, and conveniently there is a "Copy URL" button. What you will want do is connect to the web SSH that RunPod offers. You will probably start in /madapps/, in that case cd to ComfyUI, then cd to models.
On the ComfyUI screen, where it shows you what files you are missing - it also shows you what folder they go into. So let's say you need a couple of files to go into diffusion_models. Back to SSH - cd into diffusion_models. Then click the "Copy URL" button from ComfyUI. Then back in SSH all you need to know is to just add -LJO to the end of your paths. So now that you're SSHd into the diffusion_models folder, all you have to do is type curl -LJO and then paste the URL you copied from the ComfyUI button.
No quick deploy here because you need to launch/start the process via the Storage page. But then you put this: docker.io/unaico/comfypod:latest
8188: ComfyUI web interface22: SSH access
You can customize ComfyUI startup arguments by editing /workspace/madapps/comfyui_args.txt. Add one argument per line:
--max-batch-size 8
--preview-method auto
/workspace/madapps/ComfyUI: Main ComfyUI installation/workspace/madapps/comfyui_args.txt: Custom arguments file
This project is licensed under the GPLv3 License.