|
--- |
|
license: creativeml-openrail-m |
|
tags: |
|
- stable-diffusion |
|
- anime |
|
- aiart |
|
pipeline_tag: text-to-image |
|
--- |
|
|
|
**This model is trained for 5 characters from Saving 80,000 Gold in Another World for My Retirement (老後に備えて異世界で8万枚の金貨を貯めます)** |
|
|
|
Most information along with examples can be found on the civitAI model page |
|
https://civitai.com/models/17336/roukin8-character-lohaloconfullckpt-8 |
|
|
|
#### Usage |
|
|
|
There are three formats: loha, locon, and full checkpoint |
|
The first two are part of the [LyCORIS](https://github.com/KohakuBlueleaf/LyCORIS) project and can be used as lora in a1111-webui once this [extension](https://github.com/KohakuBlueleaf/a1111-sd-webui-locon) is installed |
|
|
|
You probably want to separate names by ; as this is how the model is trained. E.g. |
|
> Sabine; YamanoMitsuha; multiple girls, 2girls, bow, sliding doors, apron, dress, indoors, holding, bowtie, smile, open door, door |
|
|
|
As usual, two characters are fine, but more are difficult |
|
|
|
|
|
#### Dataset Description |
|
|
|
The dataset is prepared via the workflow https://github.com/cyber-meow/anime_screenshot_pipeline |
|
|
|
It contains 2948 roukin anime screenshots and 19279 regularization images. |
|
The anime screenshots can be downloaded from civitai. |
|
|
|
|
|
#### Training |
|
|
|
- Common: clip skip 1, resolution 512, batch size 8, on top of [Crosstyan/BPModel](https://huggingface.co/Crosstyan/BPModel) |
|
- Full checkpoint: 2.5e-6 cosine scheduler, Adam8bit, conditional dropout 0.08 |
|
- Loha/LoCon: 2e-4 constant scheduler, AdamW |
