Studio Ghibli Style LoRA - CivArchive (CivitAI Archive)

Studio Ghibli Style LoRA - offset

Studio Ghibli style LoRA

Making models can be expensive. Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕

I know there are already various Ghibli models, but with LoRA being a thing now it's time to bring this style into 2023. This is good around 1 weight for the offset version and 0.65 for the old one, on Anything v4.5 for a more authentic style, but it's also good on AbyssOrangeMix2. Triggers with ghibli style and, as you can see, it should work with any character LoRA or TI that didn't overfit on style too. Keep in mind, this is a style model not a "Ghibli characters" model, so movie characters in the examples are made using careful prompting, meaning it can reproduce similar characters, but it won't make them perfectly (unless you also use other TI/LoRA's).

4th pic by Karkaz using also his own character LoRA.

Last pic by LottaLewds.

How to use LoRA's in auto1111:

Update webui (use git pull like here or redownload it)
Copy the file to stable-diffusion-webui/models/lora
Select your LoRA like in this video
Make sure to change the weight if needed (by default it's :1)

Description

FAQ

Comments (21)

SoManyWildThingFeb 4, 2023· 1 reaction

CivitAI

Eyuuus, so excited!

treksisFeb 4, 2023· 1 reaction

CivitAI

gosh... you are so good. is this LoRA dreambooth or fine tunning? spectrums are just amazing... zelda in combo WOW

3418Feb 4, 2023· 1 reaction

CivitAI

Legendary

nolanaatamaFeb 4, 2023· 2 reactions

CivitAI

WHOA

smolspiteFeb 4, 2023

CivitAI

Is there your guide/settings for your LORA's training?

Lykon

Author

Feb 4, 2023

no, I never made a guide, nor will I. There are more competent people doing it already.
I may make a guide on how to do a good dataset tho, as that's 99% of the work.

smolspiteFeb 4, 2023

@Lykon that would be nice. Still, i don't ask for in-depth guide, just what main settings such as learning rate, epochs, batch, clip skip do you use for training?

Lykon

Author

Feb 4, 2023

just go with the default ones. It's really not important.

marson666Mar 20, 2023

@Lykon hi bro, your Lora looks so amazing, will you share your experience about making good datasets for training high quality Lora?

stygianwizard42Feb 4, 2023

CivitAI

amazing, works ok with the Ghibli grain upscale model, did you use any of the screenshots they released for artists to use ("within the scope of common sense") a couple of years ago in the model (https://www.ghibli.jp/info/013344/) ?

Lykon

Author

Feb 4, 2023

no, I just used screencaps from the movie. Making a model to imitate a style (also not 100% perfectly) should be fair use.

stygianwizard42Feb 4, 2023

@Lykon the images were released as fair use, also what do you commonly use to source images for the models

Lykon

Author

Feb 4, 2023

@stygianwizard42 fair use really depends on what you do with those images. If you sell them, that's not fair use anymore. Anyway, what I use to make the dataset really depends on what I'm trying to achieve.

NickKolokFeb 7, 2023

CivitAI

Your works look fantastic, but could you please also upload a .pt file alongside with .safetensors one? I couldn't get .safetensors working on AUTO1111: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/7172

Lykon

Author

Feb 7, 2023

How do you know that saving as pt would work? Anyway you can just checkout a previous commit hash

NickKolokFeb 7, 2023

@Lykon Well, I have an extension for WebUI that supports .pt but not .safetensors. Some guys say that it is the solution.

NickKolokFeb 7, 2023

@Lykon , however, if you share the AUTO1111 commit SHA which work for you, it will also be a great option!

Lykon

Author

Feb 7, 2023

@NickKolok I'm 148 commits behind, SHA 7a14c8ab45da8a681792a6331d48a88dd684a0a9

NickKolokFeb 9, 2023

I confirm working on 7a14c8ab45da8a681792a6331d48a88dd684a0a9 , thank you very much!

gaspachoAIFeb 8, 2023· 1 reaction

CivitAI

What "\" are for?

Lykon

Author

Feb 8, 2023· 5 reactions

in training some booru tags have "(" and ")", but those have a meaning when prompting (they're weights). The "\" in front is a way to escape them so that they're read as "(" and ")" and not weights by the generator.

LORA

SD 1.5

by Lykon

Download (Beta) View on CivitAI

Details

Downloads

103,906

Platform

CivitAI

Platform Status

Available

Created

2/4/2023

Updated

5/11/2026

Deleted

Trigger Words:

ghibli style

Files

ghibli_style_offset.safetensors

Size:

144.11 MB

SHA256:

6bb3719e90d33a9e88ea090cd4463d15ef68234e2acaac2a482186a8b7a30878