ghost
Sign In

Atla Fayon - The Rising of the Shield Hero | アトラ - 盾の勇者の成り上がり

88
469
13
Type
LoRA
Stats
469
Reviews
Published
Dec 14, 2023
Base Model
SD 1.5
Training
Steps: 5,640
Epochs: 24
Usage Tips
Clip Skip: 2
Strength: 0.8
Trigger Words
1girl, closed eyes, white hair, two-tone hair, tiger ears, tiger tail, white dress, multicolored dress, green dress, short sleeves, cross-laced, black skirt, twin braids, low twin braids, hair ornament, bracelet, black footwear
Hash
AutoV2
2DB1A85044
default creator card background decoration
SDXL Training Contest Participant
Machi's Avatar
Machi

LoRA model for Atla from The Rising of the Shield Hero.

Trained using a mixture of anime screencaps, official artworks and fan artworks.

Recommended settings

  • LoRA network weight 0.6-0.9

  • Please generate using highres fix for best effect.

For LoRA Block Weights:

  • For her default appearance, dropping OUTS blocks (1,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0) seems good.

  • For doing other outfits, or getting more freedom of poses/composition, using only OUTD blocks (1,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,0) gives more flexibility, but her age appearance will need to be prompted for

Suggested prompts

Regular Outfit (S3 anime) (Pre-vol 15)

1girl, closed eyes,
white hair, two-tone hair, long hair,
tiger ears, tiger tail,
white dress, multicolored dress, green dress, short sleeves, cross-laced, black skirt,
twin braids, low twin braids, hair ornament,
bracelet,
black footwear,

This LoRA was primarily trained on this outfit.

Known issues:

  • Her hair ornaments and bracelets sometimes double up. Sometimes this can be fixed on highres fix using Euler a.

  • The model will only sometimes put her tail below her skirt, and can often depict it too long. SD has trouble with tail placement in general, but sometimes it can help to delay the drawing of the tail by using [:tiger tail:0.2] or similar. Adding modifier tokens like "tiger tail under" or "tiger tail below" has some marginal benefit.

  • It's relatively common for the number of turns in her braids to increase from model, especially for tall images.

  • Generation of views from side and back of her clothing are less consistent, but can work (I specifically added training data to get it better at this but it's pretty limited still)

  • Her vol 15 outfit is not supported