Prompt Diffusion 🚀

This model was build via parameter-efficient QLoRA finetuning of Mistral-7B-v0.1 on a custom dataset. The goal of this project is to build a model which takes simple and short natural language inputs and generate a highly detailed and descriptive prompt for Stable Diffusion and other similar image generation models. Hence making it easier for everyone to get high quality images easily.

Note

This is the first iteration of this project and it was to check the capabilities of Mistral-7B model. As of now the model is fine-tuned mostly on the prompts which includes humans, with description of the subject and the scene. This is a work in progress and this is just a prototype. Hence, one can expect issues with response from the model.

Input Template

###Human:
generate a stable diffusion prompt for {input}

###Assistant:

Output Template

###Human:
generate a stable diffusion prompt for a model in a cafe sipping coffee

###Assistant:
A model in a cafe sipping coffee, with a warm glow illuminating her face. She wears a stylish dress and has a book in her hand. The background is a cozy cafe with dim lights.

Example prompts and responses

Example 1:

User:

a model in a cafe sipping coffee

Prompt Diffusion

A model in a cafe sipping coffee, with a warm glow illuminating her face. She wears a stylish dress and has a book in her hand. The background is a cozy cafe with dim lights.

Example 2:

User:

a man skiing

Prompt Diffusion

A man in a ski suit, with a helmet and goggles, is skiing down a snowy mountain, with the wind blowing in his face and the sun shining brightly. The man is enjoying the thrill of skiing, as he glides down the slopes, leaving behind a trail of fresh powder.

Example 3:

User:

a beautiful woman dancing in rain

Prompt Diffusion

A beautiful woman, dressed in a flowing white gown, dances gracefully in the rain. The raindrops fall gently around her, creating a magical and ethereal atmosphere. She moves with grace and elegance, her long hair flowing in the wind. The rain creates a shimmering effect on her gown, making her look like a goddess.

Example 4:

User:

a man on a beach

Prompt Diffusion

A man sits on a beach, the waves crashing against the shore. The sun is setting, casting a warm glow over the sand and water. The man is lost in thought, his mind wandering as he takes in the beauty of the scene.

Demo

This code can be run even on the free version of Google Colab. Change the runtime to GPU - T4 and run the notebook below:

Try on Colab 🚀

Basic usage

!pip install git+https://github.com/huggingface/transformers
!pip install git+https://github.com/huggingface/peft.git
!pip install torch
!pip install -q bitsandbytes accelerate

#Importing libraries

from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
import re

#Loading adapter model and merging it with base model for inferencing

torch.set_default_device('cuda')

peft_model_id = "abhishek7/Prompt_diffusion-v0.1"

config = PeftConfig.from_pretrained(peft_model_id)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    low_cpu_mem_usage=True,
    load_in_4bit=True,
    quantization_config=bnb_config,
    torch_dtype=torch.float16,
    device_map="auto"
)

model = PeftModel.from_pretrained(model, peft_model_id)

model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path, trust_remote_code=True)
tokenizer.padding_side = "right"

# Function to truncate text based on punctuation count

def truncate_text(text, max_punctuation):
    punctuation_count = 0
    truncated_text = ""
    for char in text:
        truncated_text += char
        if char in [',', '.']:
            punctuation_count += 1
            if punctuation_count >= max_punctuation:
                break
    # Replace the last comma with a full stop if the last punctuation is a comma
    if truncated_text.rstrip()[-1] == ',':
        truncated_text = truncated_text.rstrip()[:-1] + '.'

    return truncated_text


# Function to generate prompt

def generate_prompt(input, max_length, temperature):
    input_context = f'''
###Human:
generate a stable diffusion prompt for {input}

###Assistant:
'''

    inputs = tokenizer.encode(input_context, return_tensors="pt")
    outputs = model.generate(inputs, max_length=max_length, temperature=temperature, num_return_sequences=1)

    output_text = tokenizer.decode(outputs[0], skip_special_tokens = True)

    # Extract the Assistant's response using regex
    match = re.search(r'###Assistant:(.*?)(###Human:|$)', output_text, re.DOTALL)
    if match:
        assistant_response = match.group(1)
    else:
        raise ValueError("No Assistant response found")



    # Truncate the Assistant's response based on the criteria
    truncated_response = truncate_text(assistant_response, max_punctuation=10)

    return truncated_response

# Usage:

input_text = "a beautiful woman dancing in rain"

prompt = generate_prompt(input_text, 150, 0.3)

print("\nPrompt: " + prompt)

Contributing

Contributions are welcome! If you find any bugs, create an issue or submit a pull request with your proposed changes.

Acknowledgements

This model was finetuned by Abhishek Kalra on Sep 29, 2023 and is for research applications only.

mistralai/Mistral-7B-v0.1 citation

coming

Framework versions

• PEFT 0.6.0.dev0