Updated Dec/12/2023:
New version trained with Prodigy optimizer, also updated & refined dataset. Should be a whole lot less lossy.
Features HMS Plymouth from Azur Lane in her original costume & Heart-Throbbing Refreshment bikini costume, trained on animefull-pruned (legacy versions on ACertainModel) with good level of flexibility.
Usage:
Works fine on full 1.0 weight, see Trigger Words section for seperate costume prompts. (Some tags were specifically tailored/changed to prevent spillage)
Previous Update Mar/4/2023:
After a week's futile struggle with the Lion optimizer, I turned to more efficient methods such as the D-Adaption; hence the remastered 2.0 and 2.1 versions are no longer overfitted, more compatible across different checkpoints and make up only half the original file size without degrading her details.
Release v1: applying with a high weight (0.7-0.9) is recommended, going below will noticeably impact character robustness.
Release v2.0 & 2.1: Works well up to 1.0 full weight.
Prompts (Legacy ver.)
(Note that v2.0 & 2.1 takes plmtazln as trigger for her face & hairstyles)
Base: white dress, bare shoulders, half gloves, high heels, blunt bangs
Swimsuit: bikini, hair ribbon, twintails, blunt bangs
Description
FP16, dataset size 60 pics
FAQ
Comments (3)
thank you!But，why does my picture look like a mosaic?Ծ‸Ծ
It's most likely caused by omitting --no-half-vae command lines, or you have either a corrupted sd checkpoint file / applied the LoRA at too large a weight (above 1).
@Tugl Thank you!(^_^)