Combined image with WebUI

ai-dock / kohya_ss

Kohya's GUI docker images for use in GPU cloud and local environments. Includes AI-Dock base for authentication and improved user experience.

Other

10 stars 4 forks source link

Combined image with WebUI #3

Open ddpasa opened 3 months ago

ddpasa commented 3 months ago

It is nice to combine Automatic1111 WebUI together with Kohya, mainly to test the models in between training. Is this possible to merge the two docker images into a single one?

chrissy0 commented 2 months ago

I'm also looking for this

robballantyne commented 2 months ago

@ddpasa @chrissy0 I've been really busy and missed the issue. Sorry about that.

I can add this as it seems to make sense, but would Forge suffice? There seems to be more development going on there, particularly for Flux and it's otherwise A1111 compatible.

Interested in your thoughts

chrissy0 commented 2 months ago

I'm not familiar with the differences, but as long as the basic functionality is similar I'm totally fine with both. I just want to avoid renting two GPUs at the same time for a task that can easily be done on a single GPU. Thanks for your great work btw.

ddpasa commented 2 months ago

@ddpasa @chrissy0 I've been really busy and missed the issue. Sorry about that.

I can add this as it seems to make sense, but would Forge suffice? There seems to be more development going on there, particularly for Flux and it's otherwise A1111 compatible.

Interested in your thoughts

Forge makes more sense. For example forge already supports Flux but Automatic1111 is still lagging.

lutzapps commented 2 months ago

I would opt for a combination of "FORGE + ComfyUI + Kohya_ss"

Kohya_ss for training a LoRA (potentially it will be a FLUX LoRA anyway most of the times)
FORGE for testing the LoRA model and compare different LoRA epochs with S&R X/Y/Z
ComfyUI for really using your custom LoRA with FLUX

During a training you still could play or prepare a ComfyUI workflow while waiting, and all FLUX models could be shared between all 3 environments. THAT would be my wish from @robballantyne ;-)

paulpierre commented 2 weeks ago

I second this. I'm trying to run both docker-compose manifests for ComfyUI and Forge and despite setting offset ports, I'm getting errors with syncthing

supervisor_1  | [start] 2024/11/19 12:56:40 WARNING: Error opening database: resource temporarily unavailable (is another instance of Syncthing running?)

so effectively i can only run one docker-compose instance at a time.

cross-compat on the same VM would be nice or even a single docker-image.

great work btw. you're a mensch for making this for the rest of us 👏👏👏

robballantyne commented 2 weeks ago

@paulpierre pass SUPERVISOR_NO_AUTOSTART="syncthing" to one of the containers.

This variable takes a comma delimited string of services you don't want to start.

This is the first move towards separating the services over several images - Focus has always been on the all-in-one image required by some cloud platforms but it's not ideal when orchestrating locally and some services simply don't make sense on a local machine

paulpierre commented 2 weeks ago

@paulpierre pass SUPERVISOR_NO_AUTOSTART="syncthing" to one of the containers.

This variable takes a comma delimited string of services you don't want to start.

This is the first move towards separating the services over several images - Focus has always been on the all-in-one image required by some cloud platforms but it's not ideal when orchestrating locally and some services simply don't make sense on a local machine

worked 👌

for anyone else wondering, add "SUPERVISOR_NO_AUTOSTART=jupyter,syncthing" to the environment: key in docker-compose.yaml