Alita: Battle Angel DOLLBODY

Trained Locally on Kohya-Trainer: RTX 3090

4K HDR Screenshots downscaled to: 1024x900

Number of Trained Images : 468

Captioned Locally: Joy Caption

Training Parameter

Repeats: 2 (since we have a plenty of images, no need to go higher)

Epochs: 10 ( I found my sweet spot here)

Train batch size: 2 (I don't recommend going higher than 2 or 3, better to keep it at 1)

Rank: 32 ( Rank is all about details/Traits, main feature of our subject (Dimensionality of our subject to embed for the Latent space), For Alita's body a Rank of 32 is enough for me as of now. Lack of resources for training :( I would go 128k

steps: 4680 (Took about 7+Hrs to train)

Prompting

Trigger word : AL1IT0A

CFG: 1 -1.3 could work fine.

Dollbody: "The image is a high-resolution photograph depicting a woman with fair skin and straight, shoulder-length dark brown hair. She has a medium build and is dressed in a form-fitting, white bodysuit with intricate, metallic, and organic-looking patterns. The bodysuit is made of a shiny, futuristic material with a slight sheen, suggesting advanced technology. Her expression is contemplative."

[Below this prompt you can add your own]

or it would be much better if you use the joy caption texts in the dataset I provided.

Implementations

• Not only you can create Alita with this lora, but you can also merge or mix her Cyborg Berserker body , in this example with Dua Lipa by @steffangund

https://civarchive.com/user/steffangund

• I also provided the training data so you can do your own experiment & share your results as always.

DuaLipa swan song

Trailer

I'm just an insignificant Lora Creator, but a Buzz would help;)