File size: 2,681 Bytes
a982013
9e45f8d
 
 
 
 
a982013
9e45f8d
 
a982013
9e45f8d
 
 
675c8ee
 
60ddcb5
675c8ee
9e45f8d
 
 
675c8ee
9e45f8d
675c8ee
9e45f8d
 
 
 
 
675c8ee
 
 
9e45f8d
 
 
675c8ee
9e45f8d
 
 
2b70c93
9e45f8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
675c8ee
 
 
9e45f8d
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
---
language:
- en
tags:
- stable-diffusion
- text-to-image
license: bigscience-bloom-rail-1.0
inference: false

---

# waifu-diffusion - Diffusion for Weebs

waifu-diffusion is a latent text-to-image diffusion model that has been conditioned on high-quality anime images through [Textual Inversion](https://github.com/rinongal/textual_inversion).

<img src=https://cdn.discordapp.com/attachments/872361510133981234/1016022078635388979/unknown.png?3867929 width=40% height=40%>
<sub>Prompt: touhou 1girl komeiji_koishi portrait</sub>

## Model Description

The model originally used for fine-tuning is [Stable Diffusion V1-4](https://huggingface.co/CompVis/stable-diffusion-v1-4), which is a latent image diffusion model trained on [LAION2B-en](https://huggingface.co/datasets/laion/laion2B-en).

The current model is based from [Yasu Seno](https://twitter.com/naclbbr)'s [TrinArt Stable Diffusion](https://huggingface.co/naclbit/trinart_stable_diffusion) which has been fine-tuned on 30,000 high-resolution manga/anime-style images for 3.5 epochs.

With [Textual Inversion](https://github.com/rinongal/textual_inversion), the embeddings for the text encoder has been trained to align more with anime-styled images, reducing excessive prompting.

## Training Data & Annotative Prompting

The data used for Textual Inversion has come from a random sample of 25k Danbooru images, which were then filtered based on [CLIP Aesthetic Scoring](https://github.com/christophschuhmann/improved-aesthetic-predictor) where only images with an aesthetic score greater than `6.0` were used.

Captions are Danbooru-style captions.

## Downstream Uses

This model can be used for entertainment purposes and as a generative art assistant.

## Example Code

```python
import torch
from torch import autocast
from diffusers import StableDiffusionPipeline

model_id = "hakurei/waifu-diffusion"
device = "cuda"


pipe = StableDiffusionPipeline.from_pretrained(model_id, use_auth_token=True)
pipe = pipe.to(device)

prompt = "a photo of reimu hakurei. anime style"
with autocast("cuda"):
    image = pipe(prompt, guidance_scale=7.5)["sample"][0]  
    
image.save("reimu_hakurei.png")
```

## Team Members and Acknowledgements

This project would not have been possible without the incredible work by the [CompVis Researchers](https://ommer-lab.com/) and the author of the original finetuned model that this work was based upon, [Yasu Seno](https://twitter.com/naclbbr)!

Additionally, the methods presented in the [Textual Inversion](https://github.com/rinongal/textual_inversion) repo was an incredible help.

- [Anthony Mercurio](https://github.com/harubaru)
- [Salt](https://github.com/sALTaccount/)