README.md

metadata

license: creativeml-openrail-m
tags:
  - stable-diffusion
  - prompt-generator
  - distilgpt2

DistilGPT2 Stable Diffusion Model Card

DistilGPT2 Stable Diffusion is a text-to-text model used to generate creative and coherent prompts for text-to-image models, given any text. This model was finetuned on 2.03 million descriptive stable diffusion prompts from Stable Diffusion discord, Lexica.art, and (my hand-picked) Krea.ai. I filtered the hand-picked prompts based on the output results from Stable Diffusion v1.4.

Compared to other prompt generation models using GPT2, this one runs with 50% faster forwardpropagation and 40% less disk space & RAM.

PyTorch

pip install --upgrade transformers

# download DistilGPT2 Stable Diffusion if haven't already
import os
if not os.path.exists('./distil-sd-gpt2.pt'):
    import urllib.request
    print('Downloading model...')
    urllib.request.urlretrieve('https://huggingface.co/FredZhang7/distilgpt2-stable-diffusion/resolve/main/distil-sd-gpt2.pt', './distil-sd-gpt2.pt')
    print('Model downloaded.')

from transformers import GPT2Tokenizer, GPT2LMHeadModel

# load the pretrained tokenizer
tokenizer = GPT2Tokenizer.from_pretrained('distilgpt2')
tokenizer.add_special_tokens({'pad_token': '[PAD]'})
tokenizer.max_len = 512

# load the fine-tuned model
import torch
model = GPT2LMHeadModel.from_pretrained('distilgpt2')
model.load_state_dict(torch.load('distil-sd-gpt2.pt'))

# generate text using fine-tuned model
from transformers import pipeline
nlp = pipeline('text-generation', model=model, tokenizer=tokenizer)
ins = "a beautiful city"

# generate 10 samples
outs = nlp(ins, max_length=80, num_return_sequences=10)

# print the 10 samples
for i in range(len(outs)):
    outs[i] = str(outs[i]['generated_text']).replace('  ', '')
print('\033[96m' + ins + '\033[0m')
print('\033[93m' + '\n\n'.join(outs) + '\033[0m')

Example Output: