README.md

---
license: other
thumbnail: >-
  https://huggingface.co/ddPn08/subtle/resolve/main/images/main.png
tags:
  - stable-diffusion
  - text-to-image
  - safetensors
  - diffusers
inference: true
widget:
  - text: >-
      masterpiece, best quality, high quality, newest, new, exceptional, best aesthetic, anime, 1girl, short hair, black hair, messy hair, black eyes, ahoge, white shirt, outdoors, dappled sunlight, leaf, looking at viewer,
    example_title: example 1girl
  - text: >-
      masterpiece, best quality, high quality, newest, new, exceptional, best aesthetic, anime, 1boy, short hair, black hair, messy hair, black eyes, ahoge, white shirt, outdoors, dappled sunlight, leaf, looking at viewer,
    example_title: example 1boy
language:
  - en
library_name: diffusers
---

<div align="center">

![](./images/main.png)

</div>
<h1 align="center"><b>Subtly</b></h1>
<p align="center">A very subtle model. And an anime model that does not use any NAI leaks. </p>

<div align="center">

[![](https://img.shields.io/static/v1?label=FP32&message=Checkpoint&labelColor=444&color=72B76C&style=for-the-badge)](https://huggingface.co/ddPn08/subtle/resolve/main/subtly-fp32.ckpt)
[![](https://img.shields.io/static/v1?label=FP32&message=SafeTensors&labelColor=444&color=72B76C&style=for-the-badge)](https://huggingface.co/ddPn08/subtle/resolve/main/subtly-fp32.safetensors)

</div>


<br >

<div align="center">

[![](https://img.shields.io/static/v1?label=FP16&message=Checkpoint&labelColor=222&color=72B76C&style=for-the-badge)](https://huggingface.co/ddPn08/subtle/resolve/main/subtly-fp16.safetensors)
[![](https://img.shields.io/static/v1?label=FP16&message=SafeTensors&labelColor=222&color=72B76C&style=for-the-badge)](https://huggingface.co/ddPn08/subtle/resolve/main/subtly-fp16.safetensors)

</div>


<br >

# Examples
![](./images/example-1.png)

<details>
<summary>Prompt</summary>

```yaml
prompt: masterpiece, best quality, high quality, newest, new, exceptional, best aesthetic, BREAK
[anime], 1girl, black hair, short hair, messy hair, wavy mouth, sweat, nude, BREAK
simple background, looking at viewer,
negative_prompt: old, oldest, worst quality, low quality, normal quality, deleted, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry, head out of frame,
Steps: 29
Sampler: Euler a
CFG scale: 8
Seed: 3159754303
Size: 768x768
Model hash: aa0bd9fbc5
Model: subtly-fp32
Clip skip: 2
```
</details>


<br>

![](./images/example-2.png)

<details>
<summary>Prompt</summary>

```yaml
prompt: masterpiece, best quality, high quality, newest, new, exceptional, best aesthetic, BREAK
1girl, short hair, messy hair, ahoge, jitome, BREAK
outdoors, dappled sunlight, leaf, looking at viewer,
negative_prompt: old, oldest, worst quality, low quality, normal quality, deleted, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry
Steps: 29
Sampler: Euler a
CFG scale: 8
Seed: 913456778
Size: 704x704
Model hash: aa0bd9fbc5
Model: subtly-fp32
Clip skip: 2
```
</details>


<br>

![](./images/example-3.png)

<details>
<summary>Prompt</summary>

```yaml
prompt: masterpiece, best quality, high quality, newest, new, exceptional, best aesthetic, BREAK
[anime], 1girl, long hair, black hair, black eyes, open mouth, bikini, BREAK
outdoors, ocean, beach, palm tree,
negative_prompt: worst quality, low quality, normal quality, deleted, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry, old, oldest

Steps: 29
Sampler: Euler a
CFG scale: 8
Seed: 3704255935
Size: 704x704
Model hash: aa0bd9fbc5
Model: subtly-fp32
Clip skip: 2
ENSD: 31337
```
</details>


<br>

# About this
This model is trained based on Waifu Diffusion 1.5. NovelAI leakage models are not included at all. But it still won't beat NovelAI. (Subtle, isn't it?)

The Waifu Diffusion 1.5 prompt differs slightly from the NAI-based model.

For more information, please refer to this release note.

https://cafeai.notion.site/WD-1-5-Beta-Release-Notes-967d3a5ece054d07bb02cba02e8199b7

<br>

# limitation

## Realistic pictures are output.
Waifu Diffusion 1.5 learns many images from Instagram and does not output anime images by default. Therefore, you need to put `anime` in the prompt. (You may still get realistic images. Subtle.)

It can also be controlled by putting `instagram`, `real-life` in the negative prompt. However, these are not recommended because they may affect the human anatomy and the background.


<br >

## NSFW
This model is very vulnerable to NSFW. We are currently working on making it stronger to NSFW but it is not going well. (It's subtle.)

I would like to know if anyone has any knowledge about this.


<br >

# License
Subtly is released under the Fair AI Public License 1.0-SD (https://freedevproject.org/faipl-1.0-sd/).