Type | |
Stats | 413 |
Reviews | (67) |
Published | May 5, 2023 |
Base Model | |
Training | Steps: 80,000 Epochs: 8 |
Hash | AutoV2 282AF4504A |
Version Final:
A LoRA for u_u_zan's Style based on Any_Lora.
Newest larger model has a more consistent style, but this has tradeoffs:
- Worse hands
- Touhou tends to mix into character prompts and loras
- Color pallete, or backlight effects are prominent (I personally like this)
Shouldn't need any trigger words, however adding 'u_u_zan' or 'u_u_zan style', may force the style more. Works best at weight 0.8-1.0. Clip Skip 2, however it is not required to generate good images. Trained on booru tags. Good SFW.
There were no NSFW images for training, but somehow it can do it.
Tends to apply a drawing style to background.
For best results, or results similar to mine, most anime-style models should work. Additionally, it is most recommended to use badv5/EasyNegative textual inversions.
I often use variation seed, and these do not appear in some image parameters unfortunately. I encourage you to experiment with it. I use 0.05-0.15 strength.
Should work fine with other LoRAs.
Examples done using many different models. Follow prompts to attempt replication and use given negatives for better generations.
Usage:
CFG Scale: 7-9
LoRA Weight: 0.6-0.8, 1.0
Should work fine for less detailed prompts. High-res fix is highly recommended