Back to the Dawn Styles & Chars

This is mainly a style lora for Back to the Dawn official arts and pixel arts. But it can also do some major characters from game. Back to the Dawn is a prison breaking + crpg genre game, setting in a world full of anthros, where you play as a falsely accused journalist or a undercover agent/cop. It has stylized pixel arts and HD2D scenes.

Trained and tested on NoobAI-XL Eps 1.1.

Base Prompt format: {art style tag}, {character tag|species}, anthro, solo, {gender tag}, ...

Artwork Style Tags

• official art: the style of the game's steam cover art.

• pxiel \(artwork\): the escape ending pixel art. If it generates sprite size character/composition, try view point tags such as half-length portrait, or img2img with high noise on a non-pixel generation of good composition.

Other Style Tags

• Single sprite: hd pixels, sprite, white background, generate at 1024×1024.

• Sprite sequence: hd pixels, sprite, sequence, multiple {poses|angles}, white background, generate at 1024×1024. There is randomness in sprite size because different species in the game have different heights.

• In-game 2D scene: hd pixels, this style is weaken by sprites. Better up the lora strength, avoid background tags associated with deep scenery and depict floor/wall/furniture by tags.

• In-game HD2D scene: hd pixels, 3d background.

• Semi-realism: photorealism, learned from a few official fursuit photo. However, noob itself is bad at realism so don't expect much realism. If your subjects is usually depicted in very toony/anime style, the result might be strange. It's suggested to upweight this tag by 1.2 and delay this tag for 10%~20% steps with prompt editing.

Here are how each style generalize on characters from other series (legoshi from Beastars for example):

Character Tags

There are lots of characters in the game and only a few are brought into the dataset. Our two protagonists (Thomas and Bob) has clear non-pixel official art, while the others only have pixel sprites. The lora can generalize what other trained sprite-only characters look like to some extent.

Here are character tags in dataset:

• thomas fox

• bob black panther, add athletic male if needed

  • causal attire: black jacket, red shirt, jeans

• angelo wolf

  • canon attire: trilby \(hat\), white collared shirt, green vest, grey pants, black shoes

• bruce mastiff, add musclegut if needed, wears police uniform, police hat

• billy bull terrier, wears police uniform, police hat

• chris lynx

• crunchy crocodile, add muscular male if needed, he always has his towel around neck

• fenni wolf

• maggie white ferret

• beth poodle

• john tiger, add muscular male if needed

• valdimir polar bear, wears white framed eyewear, semi-rimless eyewear, has arm scar

• reed red panda, a little too young in dataset, wears glasses

• whitey rabbit

• nood raccon, his name is noodle but I have to avoid confusing with the actual food

• wesley cheetah

  • canon attire: red jacket, purple shirt, rolled up sleeves, grey pants, black boots

For prisoner uniforms in game: {red|orange|blue} prison uniform

Tips

• Negatives: worst quality, abstract background

• When using pixel \(artwor\), putting sprite in negative will give smoother shading and less pure colors.

• The pixel arts in training data are mostly "8 physical pixels = 1 logic pixel". Downscale generated image by 8 to get a pixel perfect version.