[New Extension] Full Net LoRA! LoCon model loader

[KohakuBlueleaf]

I make something for doing LoRA on Convolution Network, and make this extension for load locon model.

Extension:
https://github.com/KohakuBlueleaf/a1111-sd-webui-locon

LoCon:
https://github.com/KohakuBlueleaf/LoCon

[vladmandic]

looks interesting, but hard to judge whats the impact - can you showcase some examples with and without?

[KohakuBlueleaf]

https://civitai.com/models/14525/sakura-chiyono-o-umamusume
LoCon and LoRa trained by others

[KohakuBlueleaf]

https://civitai.com/models/14878/loconlora-yog-sothoth-depersonalization
LoCon and LoRA trained by myself.

basically just easier to keep some details (This LoRA/LoCon trained with 6imgs)

[vladmandic]

i'm sold, but need to make this code modular instead of manually patching kohya's locally checked out repo - i've created feature request for that on your repo.

[KohakuBlueleaf]

@vladmandic it looks like in Extension index, LoCon is set to added: "2022-12-04"
is this normal?

[ClashSAN]

sorry

[whitepapercg]

Is there any hope that we can see an extract LoCon of the full models that we have for LORA?

[KohakuBlueleaf]

Please check the repo...
Already have a script for it

[vladmandic]

just fyi, i'm using kohya's lora programattically, not using train_network.py script, so to incorporate locon something like this works without needing to install either lora or locon as python packages, simply by extending to pythons import path search:

lora_path = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir, 'modules', 'lora'))
sys.path.append(lora_path)
locon_path = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir, 'modules', 'locon'))
sys.path.append(locon_path)

from train_network import train

options.network_module = 'locon.locon_kohya'
train(options)

[lsaa]

can the locon loader be added to the default lora extension? or is it too early / experimental to let in

[KohakuBlueleaf]

I think it's too early
I will add lot more thing in locon (will migrate to another project for further development) and also in the ext

[2blackbar]

so the difference is that it trains resolution block? ill try it out soon on a likeness of a person cause thats like the only thing it gives me issues now

[KohakuBlueleaf]

And I use cp-decomposition on Convolution layer now.
I also introduced Hadamard product representation for parameter-efficient finetuning.
You can check my repo for more information
https://github.com/KohakuBlueleaf/LyCORIS

[KohakuBlueleaf]

@ClashSAN I made a new version extension for my LyCORIS model
https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris

It is stand alone one using ExtraNetwork api (or hook?)
it is not hijack the built-in lora system so will not be conflict with other lora extension

I think this is better implementation

[KohakuBlueleaf]

Is it ok to put this new extension into extension index?