Aayla Secura | Star Wars

This is my first LoRA. Feel free to leave some feedback and ideas for improvement.

Genereal training information

Based on 30 images cropped to 768 x 768.

Manually refined the captions from BLIP afterwards.

Run with 100 steps per image totals in 3000 steps.

Trained on SD1.5

GTX1080 ~ 4h

I did some basic testing, the following settings can be tried out as a baseline:

• Model: SD1.5, however, other models are working well too (Mostly tested with The Ally's Mix)

• CFG: 6-9, mostly used 7

• Steps: 22, 30, 35 worked well with 2M Karras and DDIM

• 768 by 768 will probably yield best results as it was trained on this dimensions

• LoRA weight: 0.3 [very far away from real Aayla but works sometimes]; 0.6 [very nice middleground for getting some batches for testing]; 0.8 [also worked very nice]; 0.9 [wouldn't go higher as the results turn out weirdly]

Styles

• Realistic: (ultra realistic:1.1) [example image 1 and 2]

• Cartoon: (cartoon:1.2) [weight 1.5 really exaggerates the cartoon look, therfore face details have to be sacrificed; example image 3, 4 with latent upscaling]

• Drawing: drawing:1.1 of [pretty similar to cartoon, I personally favor cartoon]

Promts that should trigger additional training data:

• wearing star wars costume

• wearing star wars outfit

• leather headpiece

• blue light saber (put this on the negative promts if you don't want blue lights everywhere)

*most example images where upscaled with denoising around 0.55 (2 times, latent)