PixAI Membership
PixAI
PixAI
Generate
Sign in
Home
Generate
Models
Recipe
NEW
Studio
Contest
Mio.2

Conversational creation

AI Tools
Guide

User LoRA

XL

Alice Thymefield (Zenless Zone Zero)

#Zenless Zone Zero#Alice Thymefield
HomeModelsMio.2Profile

AI Tools

  • OC Maker
  • AI Anime Generator
  • VTuber Maker
  • Sonic OC Maker
  • Hazbin Hotel OC Maker
  • Anime Character Creator
  • AI Figure Generator
  • Genshin OC Maker
  • All AI Tools

Discover

  • Trending Tags
  • Ranking
  • Model Market
  • Contest
  • News

About

  • About PixAI
  • Guide
  • PixAI Blog
  • Tsubaki.2
  • Meet Mio, Our Mascot
  • Content Rules

Pricing & Help

  • Membership
  • Buy Credits
  • Contact

Mobile App

  • Get the PixAI app
  • App Store
  • Google Play
© 2026 PixAI
  • Terms of Service
  • Privacy Policy
  • Copyright Policy
  • Attributions
  • Specified Commercial Transactions Law
Alice Thymefield (Zenless Zone Zero) - AI Model cover image
Uploaded at
Jul 8, 2025, 9:29 PM
Uses
32
Reviews
No reviews
Trigger Words

zzzatf

Permissions
  • Allow image generation & sharing
  • Allow users to download this model
  • Allow commercial use of generated images

Description

Wanted to try training a character LoRA, so here's Alice Thymefield, upcoming S rank from Zenless Zone Zero.The keyword is "zzzatf." There were already some noticeable associations that Illustrious has with "alice_thymefield," so I picked an acronym for "zenless zone zero alice thyme-field." For simplicity, that's the only tag associated with the character, so the outfit might vary a little bit, but it can usually be recovered by gently nudging the model in the right direction with terms like "skirt."This LoRA was trained on Illustrious-XL-v1.0, but I've found it works best with models like Hassaku XL Illustrious and Unholy Desire Mix.The LoRA was trained on a dataset of 78 manually selected images limited to a training resolution of 1024,1024 and run for 40 epochs and 1160 steps in total with a batch size of 4, with 29 steps per epoch (I don't know where kohya gets that epoch step count from, it's about 10 higher than the naive calculation of 19.5). The learning rate for the was 10e-4 for the unet and 5e-5 for the text encoder. The AdamW optimizer was used with a cosine scheduler.

Comments