|
--- |
|
language: |
|
- en |
|
tags: |
|
- stable-diffusion |
|
- text-to-image |
|
license: bigscience-bloom-rail-1.0 |
|
inference: false |
|
thumbnail: "https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/449d42a8-28c5-44da-afd7-28d7e29a264c/public" |
|
--- |
|
|
|
[DRAFT] |
|
|
|
# pony-diffusion-v4 - "same. but different" edition |
|
|
|
pony-diffusion is a latent text-to-image diffusion model that has been conditioned on high-quality pony, furry and other non photorealistic SFW and NSFW images through fine-tuning. |
|
|
|
**WARNING:** This model is capable of producing NSFW content so it's recommended to use 'safe' tag in prompt in combination with negative prompt for image features you may want to suppress (i.e. nudity). |
|
|
|
**Despite its name, this model is capable of producing wide range of furry and cartoon images as side effect of improving data diversity (with exception of anime stlyes, for which Waifu Diffusion is much stronger choice).** |
|
|
|
Special thanks to [Waifu-Diffusion](https://huggingface.co/hakurei/waifu-diffusion) for providing finetuning expertise and advising through the process, without their help this project would not exist. |
|
|
|
[](https://colab.research.google.com/drive/1rl-R39aKp42RS6sRsM9cVC2elBFugdDd?usp=sharing) |
|
|
|
**Please join PurpleSmartAI Discord to use this model with our free SD bot and get early access to pony-diffusion-v4.** |
|
|
|
[](https://discord.gg/pYsdjMfu3q) |
|
|
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/449d42a8-28c5-44da-afd7-28d7e29a264c/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/d72e5089-3857-417e-87a2-0fa59111022d/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/3bf767d5-6770-499b-bac4-353e536b4b7d/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/740ab221-a882-4ad9-bf46-fd3d60c503b9/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/e9700843-24fe-48ef-981b-5171c60f69c4/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/f70fde75-f7ba-4c78-8229-3e39314aaf43/public width=25% height=25%> |
|
<img src=https://imagedelivery.net/_wFNZAzgWNWPmneM1cyjcw/artifact/422c26b6-38fa-4e0b-9c33-cffe4dc6f879/public width=25% height=25%> |
|
|
|
You can see more samples at [PurpleSmartAI](https://purplesmart.ai/collection/top?nsfw=0&page=1&gen_type=txt2img&model=5&order=created_desc) |
|
|
|
[PyTorch Model(Use this with Automatic1111 or other SD UIs)](https://mega.nz/file/9S0iSLpC#NqTvjJIzKoT2ScP9F5tFj7J3p5DhpLbIKiR70hMXqQ4) |
|
|
|
## Model Description |
|
|
|
The model originally used for fine-tuning is [Stable Diffusion V1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5), which is a latent image diffusion model trained on [LAION2B-en](https://huggingface.co/datasets/laion/laion2B-en). |
|
|
|
This particular checkpoint has been fine-tuned with a learning rate of 5.0e-6 for 20 epochs on approximately 1.7M pony, furry and other cartoon text-image pairs (using metadata from derpibooru, e621 and danbooru). |
|
|
|
## Improvements over previous models |
|
|
|
### Better disentanglement of tag based prompts |
|
Aka ["using Hidden States of CLIP’s Penultimate Layer"](https://blog.novelai.net/novelai-improvements-on-stable-diffusion-e10d38db82ac#:~:text=Using%20Hidden%20States%20of%20CLIP%E2%80%99s%20Penultimate%20Layer), a technique adopted by SD2 which should lead to generally higher quality and more tag driven outputs. |
|
In our experiments using penultimate CLIP was not always the best choice so trying both CLIP skip of 1 and 2 is recommended. |
|
|
|
### Better support for non square images and increase of default resolution to 768px |
|
This should allow you to generate full body image with resolution 512x768px without triggering "double head" glitches. |
|
|
|
### Removed SFM/3D bias |
|
V2 model demostrated bias toward 3d/sfw visual styles so special care has been applied to restrict exposure to 3d/sfm images during training. |
|
|
|
### Improver data diversity |
|
We ran multiple experiments finetuning models on 300k pony only and 600k pony only datasets, the resulting models demostrated worse quality for pony specific data. We concluded that despite more complicated prompting (and lack of pony bias by default), inclusion of large amount of non pony specific highly ranked non photorealistic images generally has strong positive efect. |
|
|
|
## License |
|
|
|
This model is open access and available to all, with a CreativeML OpenRAIL-M license further specifying rights and usage. |
|
The CreativeML OpenRAIL License specifies: |
|
|
|
1. You can't use the model to deliberately produce nor share illegal or harmful outputs or content |
|
2. The authors claims no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the license |
|
3. You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully) |
|
[Please read the full license here](https://huggingface.co/spaces/CompVis/stable-diffusion-license) |
|
|
|
## 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, DDIMScheduler |
|
model_id = "AstraliteHeart/pony-diffusion-v3" |
|
device = "cuda" |
|
pipe = StableDiffusionPipeline.from_pretrained( |
|
model_id, |
|
torch_dtype=torch.float16, |
|
revision="fp16", |
|
scheduler=DDIMScheduler( |
|
beta_start=0.00085, |
|
beta_end=0.012, |
|
beta_schedule="scaled_linear", |
|
clip_sample=False, |
|
set_alpha_to_one=False, |
|
), |
|
) |
|
pipe = pipe.to(device) |
|
prompt = "pinkie pie anthro portrait wedding dress veil intricate highly detailed digital painting artstation concept art smooth sharp focus illustration Unreal Engine 5 8K" |
|
with autocast("cuda"): |
|
image = pipe(prompt, guidance_scale=7.5)["sample"][0] |
|
|
|
image.save("cute_poner.png") |
|
``` |
|
|
|
## Team Members and Acknowledgements |
|
|
|
This project would not have been possible without the incredible work by the [CompVis Researchers](https://ommer-lab.com/). |
|
|
|
- [Waifu-Diffusion for helping with finetuning and providing starting checkpoint](https://huggingface.co/hakurei/waifu-diffusion) |
|
|
|
In order to reach us, you can join our [Discord server](https://discord.gg/WG78ZbSB). |
