This LoRA was built to perfectly replicate the Nijijourney style.
The Dataset contains around 300-320 images (depending on the base model).
Helpful Tags:
warm theme(Highly recommended)cold themeEuropeanAsianMatureYoung
(All of these tags were used when building the dataset to make it as versatile as possible.)
Recommended Weight:
If it's your main/only style LoRA: 0.8 – 1.2
If you're mixing it with other style LoRAs: 0.6 – 0.8
Generation Tips:
The lower the CFG, the better. I recommend keeping your CFG between 2 and 5 (3.5 seems to be the sweet spot for a lot of checkpoints).
You'll get the best results with samplers that pull out fine details (like DPM++ 2M).
If you want it to stick to the style as closely as possible, use NoobAI-based checkpoints that actually understand the
oil painting \(medium\)tag. Good examples are NoobAI EPS 1.1, Chenkin Noob XL, and similar models.
Description
The dataset has been expanded to almost 400 images and further cleaned up.
The style is more consistent now, but there's a slight overfitting toward darker skin tones, so I'd recommend using the pale skin tag more often.
6000 steps
4 batch
UNET_LR 0.0001
TE_LR 0.00001
Optimizer: AdamW8bit
Training model: Illustrious V2.0 Stable