app.py

import spaces # beginn
import torch.multiprocessing as mp
import torch
import os
import re
import random
from collections import deque
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
import gradio as gr
from accelerate import Accelerator

# Check if the start method has already been set
if mp.get_start_method(allow_none=True) != 'spawn':
    mp.set_start_method('spawn')

# Instantiate the Accelerator
accelerator = Accelerator()

dtype = torch.bfloat16

# Set environment variables for local path
os.environ['FLUX_DEV'] = '.'
os.environ['AE'] = '.'

# Seed words pool
seed_words = []

used_words = set()

# Queue to store parsed descriptions
parsed_descriptions_queue = deque()

# Usage limits
MAX_DESCRIPTIONS = 30
MAX_IMAGES = 12

def initialize_diffusers():
    from optimum.quanto import freeze, qfloat8, quantize
    from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL
    from diffusers.models.transformers.transformer_flux import FluxTransformer2DModel
    from diffusers.pipelines.flux.pipeline_flux import FluxPipeline
    from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast

    bfl_repo = 'black-forest-labs/FLUX.1-schnell'
    revision = 'refs/pr/1'

    scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(bfl_repo, subfolder='scheduler', revision=revision)
    text_encoder = CLIPTextModel.from_pretrained('openai/clip-vit-large-patch14', torch_dtype=dtype)
    tokenizer = CLIPTokenizer.from_pretrained('openai/clip-vit-large-patch14', torch_dtype=dtype)
    text_encoder_2 = T5EncoderModel.from_pretrained(bfl_repo, subfolder='text_encoder_2', torch_dtype=dtype, revision=revision)
    tokenizer_2 = T5TokenizerFast.from_pretrained(bfl_repo, subfolder='tokenizer_2', torch_dtype=dtype, revision=revision)
    vae = AutoencoderKL.from_pretrained(bfl_repo, subfolder='vae', torch_dtype=dtype, revision=revision)
    transformer = FluxTransformer2DModel.from_pretrained(bfl_repo, subfolder='transformer', torch_dtype=dtype, revision=revision)

    quantize(transformer, weights=qfloat8)
    freeze(transformer)
    quantize(text_encoder_2, weights=qfloat8)
    freeze(text_encoder_2)

    pipe = FluxPipeline(
        scheduler=scheduler,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        text_encoder_2=None,
        tokenizer_2=tokenizer_2,
        vae=vae,
        transformer=None,
    )
    pipe.text_encoder_2 = text_encoder_2
    pipe.transformer = transformer
    pipe.enable_model_cpu_offload()

    return pipe

def generate_description_prompt(subject, user_prompt, text_generator):
    prompt = f"write concise vivid visual description enclosed in brackets like [ <description> ] less than 100 words of {user_prompt} different from {subject}. "
    try:
        generated_text = text_generator(prompt, max_length=160, num_return_sequences=1, truncation=True)[0]['generated_text']
        generated_description = re.sub(rf'{re.escape(prompt)}\s*', '', generated_text).strip()  # Remove the prompt from the generated text
        return generated_description if generated_description else None
    except Exception as e:
        print(f"Error generating description for subject '{subject}': {e}")
        return None

def parse_descriptions(text):
    descriptions = re.findall(r'\[([^\[\]]+)\]', text)
    descriptions = [desc.strip() for desc in descriptions if len(desc.split()) >= 3]
    return descriptions

@spaces.GPU
def generate_descriptions(user_prompt, seed_words_input, batch_size=100, max_iterations=50):
    descriptions = []
    description_queue = deque()
    iteration_count = 0

    print("Initializing the text generation pipeline with 16-bit precision...")
    model_name = 'meta-llama/Meta-Llama-3.1-8B-Instruct'
    model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map='auto')
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    text_generator = pipeline('text-generation', model=model, tokenizer=tokenizer)
    print("Text generation pipeline initialized with 16-bit precision.")

    seed_words.extend(re.findall(r'"(.*?)"', seed_words_input))

    while iteration_count < max_iterations and len(parsed_descriptions_queue) < MAX_DESCRIPTIONS:
        available_subjects = [word for word in seed_words if word not in used_words]
        if not available_subjects:
            print("No more available subjects to use.")
            break

        subject = random.choice(available_subjects)
        generated_description = generate_description_prompt(subject, user_prompt, text_generator)
        
        if generated_description:
            clean_description = generated_description.encode('ascii', 'ignore').decode('ascii')
            description_queue.append({'subject': subject, 'description': clean_description})

            print(f"Generated description for subject '{subject}': {clean_description}")

            used_words.add(subject)
            seed_words.append(clean_description)

            if iteration_count % 3 == 0:
                parsed_descriptions = parse_descriptions(clean_description)
                parsed_descriptions_queue.extend(parsed_descriptions)

        iteration_count += 1

    return list(parsed_descriptions_queue)

@spaces.GPU(duration=120)
def generate_images(parsed_descriptions, pipe):
    if len(parsed_descriptions) < MAX_IMAGES:
        prompts = parsed_descriptions
    else:
        prompts = [parsed_descriptions.pop(0) for _ in range(MAX_IMAGES)]

    images = []
    for prompt in prompts:
        images.extend(pipe(prompt, num_images=1).images)

    return images

def combined_function(user_prompt, seed_words_input):
    parsed_descriptions = generate_descriptions(user_prompt, seed_words_input)
    pipe = initialize_diffusers()
    images = generate_images(parsed_descriptions, pipe)
    return images

if __name__ == '__main__':
    torch.cuda.init()

    interface = gr.Interface(
        fn=combined_function,
        inputs=[gr.Textbox(lines=2, placeholder="Enter a prompt for descriptions..."), gr.Textbox(lines=2, placeholder='Enter seed words in quotes, e.g., "cat", "dog", "sunset"...')],
        outputs=gr.Gallery()
    )

    interface.launch()