Harold & Willy - Mercury Abbey (NoobAI)

Characters and styles from a cute furry themed adventure puzzle game Mercury Abbey: Harold Graywool the novelist and his nephew Willy Fabre Highland visited the resort for vacation, and revealed a surreal conspiracy during their stay.

Trained and tested on NoobAI-XL Epsilon-pred 1.0-Version. Should be applicable on closely related merges and mixes. For triggers, see prompting section below.

Prompting

• General

  • Use quality tags as usual: masterpiece, best quality in positive, worst quality, normal quality in negative.

• Harold

  • Base trigger: maharold, wolf, solo, anthro, male

  • Optional traits: eyebrow scar, green eyes, x scar on chest

  • Novelist Outfits: grey shirt and brown vest, amber necklace, rolled up sleeves, belt, black pants

  • Hotel Outfits: bathrobe, blue robe, rolled up sleeves,

• Willy (PLEASE DON'T USE HIM FOR NSFW STUFF)

  • Base trigger: mawilly, domestic dog, solo, anthro, male, young

  • Optional traits: blue eyes

  • Outfits: scout uniform, cardboard dinosaur costume on head

• Other tips

  • Almost all data are from the game, this means the game's portrait style is branded on the lora. To make other artist tags take effect, emphasize artist tags, use lora control or lower lora weight to ~0.75.

  • Many other characters' portraits are included in the dataset, but they aren't given trigger words and are diluted to spare more training steps for Harold and Willy. You can try to prompt different characters with descriptive tags. For Jampo, try: dhole, solo, anthro, male, orange fur, tan countershading, short ponytail, red hair, white shirt and blue vest, red pocket square, blue pants, sapphire necklace.

  • I also tried to train a separate ma pixel art tag for the shaded pixel style in game. It doesn't work well sadly. However, using pixel \(artwork\) yields interesting results.

About Training

Batch size 4, 2800/4 steps. Unet LR 1e-4, TE LR 5e-5, scale LR with batch size. AdamW optimizer with cosine schedule. 16 net dim with 16 net alpha.

The game could have used more expressions for major charaters.