My first attempt at Lora training, any feedback is very appreciated!
A weight of 0.9-1.1 is recommended, though generally I believe that 1.0 gives the most consistent results (by a very small margin).
The trigger prompt is cheems_perona_v3.
Adding "red capelet, pink bowtie, striped pantyhose, crop top, red skirt and red boots" to the positive prompts will aid in consistency of her pre timeskip outfit, though her eyes are not anime accurate.
If you want to see her with her parasol, you can achieve that with the prompt red umbrella.
Since this is a first pass/first try with the dataset ive precured, there is a bit of noise with the hair and some other elements - these can usually be washed out with the prompt simple background or some inpainting.
You can also add her ghosts with the ghost tag.
The examples shown above were generated with the hires fix function, with these settings:
Denoising strength: 0.6,
Clip skip: 2,
Hires upscale: 2,
Hires steps: 30,
Hires upscaler: Latent
Look forward to improved versions of this lora in the future, and give me some more suggestions of under-loved One Piece characters!
