THIS MODEL HAS BEEN ARCHIVED AS IT NO LONGER CONFORMS TO MY CURRENT QUALITY STANDARDS!

See here for an extensive article about my entire training workflow and configuration if you want to train your own SDXL 1.0 LoRa's and achieve results similar as me!

Recommended weight: 1.0 - or lower when you combine it with other LoRa's

Recommended model and VAE: SDXL 1.0 with 0.9 vae

Recommended settings: 35 steps | DPM++ 2M Karras | CFG 7

Recommended negative prompt: text, watermark, signature, abstract, burned, instagram, low-quality, low-res, action figure, child drawing, jpeg, low contrast, glossy, puppet, cliché, stock photo, magazine cover

Description

A LoRa for SDXL 1.0 based on the artstyle and some of the characters of primarily the 1980s-early 2000s Japanese animated movies by Studio Ghibli. The character LoRa's work very well to reproduce the characters in their native artstyles, but may have trouble with outfit and style flexibility due to dataset limitations.

The sample images for the characters portrayed in the Ghibli artstyle were done using both the character LoRa at 1.0 strength and the Ghibli artstyle LoRa at 0.6 strength.

Trigger words

• Artstyle: "ghst artstyle"

• San: "s with r hairstyle wearing w outfit"

  • with fur coat

  • with crystal necklace

  • with teeth necklace

• Kiki: "v with f hairstyle wearing h outfit"

• Nausicaä: "ncpp with srhh hairstyle wearing bmcc outfit"

Further usage notes

For the character LoRa's it can help sometimes to negative prompt "anime, cartoon" to force a different artstyle. It can also help to add to the positive prompt words like "dslr" for photo prompts or "realistic" for realistic digital art.

If you notice a lack of outfit likeness, emphasizing that part of the prompt can help, too. Example:

"full-length full-body screencap in ghst artstyle of (ncpp with srhh hairstyle wearing bmcc outfit:1.3)"

Generally the characters do really well at portraying the characters in their native artstyles and in their native outfits, but it can struggle a lot with portraying them in different artstyles or outfits. Unfortunately that is hard to overcome through training because of limited datasets.

Changelog

v2.0 (Nausicaä character)

• much reduced overtraining, much improved flexibility and diversity, at the cost of some likeness

v2.0 (Ghibli artstyle)

• much reduced overtraining, much improved flexibility and diversity, at the cost of some likeness

v1.0

• Initial release

Additional Info

Credit is not mandatory but I would still appreciate credit when posting images using these LoRa's on social media! Feel free to @ my Twitter and/or Instagram accounts too (@ai_characters)! I would love to know when it is used!

I share all of my model's and training knowledge for free, but donations to my Ko-Fi would be greatly appreciated! Testing so much and training so many model's costs a lot of money (and time).

Feel free to also try out my RVC2 AI voice models which you can find here!

Stats (v2.0)

Dataset:

• Ghibli artstyle: 131x screencap

• Nausicaä: 17x screencap, 3x fanart, 9x cosplay photo

Epochs chosen for the final models:

• Ghibli artstyle: 10

• Nausicaä: 24

Learning rate:

• 3e-5 (Ghibli artstyle)

• 8e-5 (Characters)

For the configuration used to train this model see my guide which I linked in the beginning of this post.

Stats (v1.0)

Datasets:

• San: 23x screencap, 3x cosplay photo, 7x fanart

• Kiki: 26x screencap, 3x cosplay photo, 3x fanart

Epochs chosen for the final models:

• San: 70

• Kiki: 70