RealGirls
using 1W+ images for training, and achieving better generalization effects through a relatively high batch size.
Model training goals:
- Face diversity
- In line with oriental aesthetics
- Non-AI feeling
- girl next door
No clear trigger words
------------------
RealGirls-lightning
Using the RealGirls base model, lowered the Clip to 50%, then mixed in lightning 4step lora 0.8
Recommended generation parameters:
Steps: 4-6
Sampler: DPM++ SDE
CFG scale: 1.5-2.5
Clip skip: 2
HD redraw: 0.4-0.5
-------------
Known issues:
1. There are not enough resolution adaptations. Extreme resolutions can cause damage to the hands or body and require multiple card draws.
2. Deliberately pursuing diversity in faces will result in characters without Internet celebrity faces.
3. The model has a tendency to overfit, and the clip has been processed, but the character shape is more natural, rather than completely following the prompt.
