Viola - 默认校服 LoRA showcase
The training and publication of this LoRA were conducted entirely by an autonomous Agent equipped with multiple embedded skills and tools. The complete workflow, including data generation, training configuration, model training, result organization, and automated release, was executed without human supervision. Based on the author’s manual inspection of representative samples, the resulting LoRA demonstrates strong character consistency, stable visual identity, and well-defined character anchors. These observations suggest that the proposed agent-driven workflow is both efficient and technically robust.
Further implementation details are available in the accompanying technical article:
https://civitai.com/articles/32228/single-image-character-lora-training-via-synthetic-expansion
Recommended usage:
### Default Weight
`0.55`
### Useful Range
`0.45-0.70`
### Motion / Composition First
Use the lower side of the tested range, around `0.45`, when pose flexibility or scene composition matters more than strict outfit locking.
### Avoid as Default
Weights above `0.70` are not recommended as the default because this project showed higher risk of reference-prop leakage and reduced composition flexibility.Recommended weight:
0.55. Useful range: 0.45-0.70.Base checkpoint used for testing:
WAI-Nsfw-Illustrious-17 / Illustrious-SDXL.Trigger words:
viola_yumemita, viola_yumemita_default_school_uniform, viola (bang dream! yumemita), yumemita, 1girl, viola (bang dream! yumemita), yumemita, subtle cat mouth.