search

Acly / krita-ai-diffusion

Streamlined interface for generating images with AI in Krita. Inpaint and outpaint with optional text prompt, no tweaking required.
https://www.interstice.cloud
GNU General Public License v3.0
6.49k stars 313 forks source link

Docker image aclysia/sd-comfyui-krita is not up to date #818

Closed shahrin014 closed 1 month ago

shahrin014 commented 3 months ago

I notice the docker image in https://hub.docker.com/r/aclysia/sd-comfyui-krita/tags is v1.16.1

The latest version is v1.18.1

I tried to automate this using github actions but the free runner runs out of space.

Perhaps we can automate the docker build in a runner hosted on a cloud service?

Acly commented 3 months ago

It's possible, but the size is an issue regardless. The docker image is bloated with so many optional models, making it less useful to deploy on demand. Latest is 49 GB.

So I've focused more on the permanently hosted solution. I know it's not for everyone since you can't customize. But if you need custom models you might be best of building your own image anyway. In that case it makes more sense for me to publish an image with only minimal requirements?

moiri-gamboni commented 3 months ago

The permanently hosted solution you're talking about is Interstice? In any case, personally I would appreciate even a minimal image that's up to date, since I saw there were a lot of model changes in a recent update. But I'm wondering if it would not make more sense to have an image without the models altogether and have them downloaded in the post-start script?

Acly commented 3 months ago

The permanently hosted solution you're talking about is Interstice?

Yes

I uploaded an up-to-date image now, tag is v1.20.0 (next version) - but it should also work with v1.19.0. Only tested it local so far.

But I'm wondering if it would not make more sense to have an image without the models altogether and have them downloaded in the post-start script?

That would be an improvement I think. It means there's few more things that can go wrong since it doesn't all come in one pre-built blob. But it would make building the image much easier and reduce duplication.

Adding a copy of this repo to the docker image and running download_models.py should work. Maybe it's possible to forward arguments to the script by editing the "Container start command" in the runpod template? That way people could pass for example --no-sd15 and considerably decrease download size if they're not interested in SD1.5.

Acly commented 1 month ago

aclysia/sd-comfyui-krita:v1.22.0 is the first "barebone" image which downloads models after start-up.

https://github.com/Acly/krita-ai-diffusion/wiki/Cloud-GPU#-customization