82 lines
3.0 KiB
Markdown
82 lines
3.0 KiB
Markdown
---
|
||
license: creativeml-openrail-m
|
||
language:
|
||
- en
|
||
widget:
|
||
- text: 1girl, fate
|
||
- text: 1boy, league of
|
||
- text: 1girl, genshin
|
||
- text: 1boy, national basketball association
|
||
- text: 1girl, spy x
|
||
- text: 1girl, absurdres
|
||
tags:
|
||
- stable-diffusion
|
||
- anime
|
||
- anything-v4
|
||
- art
|
||
- arxiv:2210.14140
|
||
datasets:
|
||
- FredZhang7/anime-prompts-180K
|
||
---
|
||
|
||
## Fast Anime PromptGen
|
||
|
||
This model was trained on a dataset of **80,000** safe anime prompts for 3 epochs. I fetched the prompts from the [Safebooru API endpoint](https://safebooru.donmai.us/posts/random.json), but only accepted unique prompts with **up_score ≥ 8** and without any [blacklisted tags](./blacklist.txt).
|
||
I didn't release the V1 model because it often generated gibberish prompts. After trying all means to correct that behavior, I eventually figured that the cause of the gibberish prompts is not from the pipeline params, model structure or training duration, but rather from the random usernames in the training data.
|
||
Here's the complete [prompt preprocessing algorithm](./preprocess.py).
|
||
|
||
|
||
## Text-to-image Examples
|
||
|
||
Prefix *1girl* | [Generated *1girl* prompts](./anime_girl_settings.txt) | Model *Anything V4*
|
||
|
||
|
||
|
||
Prefix *1boy* | [Generated *1boy* prompts](./anime_boy_settings.txt) | Model *Anything V4*
|
||
|
||
|
||
|
||
## Contrastive Search
|
||
```
|
||
pip install --upgrade transformers
|
||
```
|
||
```python
|
||
import torch
|
||
from transformers import GPT2Tokenizer, GPT2LMHeadModel, pipeline
|
||
tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
|
||
tokenizer.add_special_tokens({'pad_token': '[PAD]'})
|
||
model = GPT2LMHeadModel.from_pretrained('FredZhang7/anime-anything-promptgen-v2')
|
||
|
||
prompt = r'1girl, genshin'
|
||
|
||
# generate text using fine-tuned model
|
||
nlp = pipeline('text-generation', model=model, tokenizer=tokenizer)
|
||
|
||
# generate 10 samples using contrastive search
|
||
outs = nlp(prompt, max_length=76, num_return_sequences=10, do_sample=True, repetition_penalty=1.2, temperature=0.7, top_k=4, early_stopping=True)
|
||
|
||
print('\nInput:\n' + 100 * '-')
|
||
print('\033[96m' + prompt + '\033[0m')
|
||
print('\nOutput:\n' + 100 * '-')
|
||
for i in range(len(outs)):
|
||
# remove trailing commas and double spaces
|
||
outs[i] = str(outs[i]['generated_text']).replace(' ', '').rstrip(',')
|
||
print('\033[92m' + '\n\n'.join(outs) + '\033[0m\n')
|
||
```
|
||
|
||
Output Example:
|
||
|
||
|
||
|
||
Please see [Fast GPT PromptGen](https://huggingface.co/FredZhang7/distilgpt2-stable-diffusion-v2) for more info on the pipeline parameters.
|
||
|
||
|
||
## Awesome Tips
|
||
|
||
- If you feel like a generated anime character doesn't show emotions, try emoticons like `;o`, `:o`, `;p`, `:d`, `:p`, and `;d` in the prompt.
|
||
I also use `happy smirk`, `happy smile`, `laughing closed eyes`, etc. to make the characters more lively and expressive.
|
||
|
||
- Adding `absurdres`, instead of `highres` and `masterpiece`, to a prompt can drastically increase the sharpness and resolution of a generated image.
|
||
|
||
## Danbooru
|
||
[Link to the Danbooru version](https://huggingface.co/FredZhang7/danbooru-tag-generator) |