|
--- |
|
license: mit |
|
language: |
|
- en |
|
base_model: |
|
- Laxhar/sdxl_noob |
|
pipeline_tag: text-to-image |
|
tags: |
|
- art |
|
datasets: |
|
- v2ray/mike-no-hito |
|
library_name: peft |
|
--- |
|
# NoobAI XL LoRA Mike No Hito |
|
This is a LoRA for the [v1.1 version of the NoobAI XL model](https://civitai.com/models/833294?modelVersionId=1116447). |
|
|
|
The dataset [v2ray/mike-no-hito](https://huggingface.co/datasets/v2ray/mike-no-hito) used to train this LoRA is scraped using [LagPixelLOL/mitgw](https://github.com/LagPixelLOL/mitgw), containing a total of 172 images. |
|
|
|
Big thanks to the artist for the very cute catgirls :3, you can find the artist on X (Twitter) with ID [@doremifaso64](https://x.com/doremifaso64). |
|
|
|
To use this LoRA, you can use the trigger word `mike no hito`. The artist is a furry artist so this will have a chance to make a character into a furry unprompted, if you encounter this, simply add `furry` to the negative prompt. |
|
|
|
This LoRA is trained using [kohya-ss/sd-scripts](https://github.com/kohya-ss/sd-scripts), with rank 32, alpha 16, learning rate 1e-4, for 512 epochs with a total of 9728 steps, using an A100-80GB, took approximately 14 hours. |
|
|
|
If you have any questions, suggestions, or just want to talk to me, you can add me on Discord with ID [@v2ray](https://discord.gg/r4Wj97nZ). |
|
|
|
## Changelog |
|
- **v2.2**: This version is trained with rank 32, for 512 epochs (384 more epochs compared to v2.1), I didn't expect it to work but it actually works pretty well. If you think it's bad you can always use an older version. |
|
- **v2.1**: This version is trained with 3x3 Conv2d layers targeted, for 128 epochs (64 more epochs compared to v2.0). |
|
- **v2.0**: This version added a lot more training data, which are hand picked, cleaned, and deduped by me, mostly by hand. It has improved non-furry ability thanks to generalization, and overall better quality. |
|
|
|
## Examples |
|
 |
|
 |
|
 |
