app.py

import os
import asyncio
import time
from generate_prompts import generate_prompt
from diffusers import AutoPipelineForText2Image
from io import BytesIO
import gradio as gr
import ray

ray.init()

@ray.remote
class ModelActor:
    def __init__(self):
        """
        Initializes the ModelActor class and loads the text-to-image model.
        """
        self.model = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo")

    async def generate_image(self, prompt, prompt_name):
        """
        Generates an image based on the provided prompt.
        Parameters:
            - prompt (str): The input text for image generation.
            - prompt_name (str): A name for the prompt, used for logging.
        Returns:
            bytes: The generated image data in bytes format, or None if generation fails.
        """
        start_time = time.time()
        process_id = os.getpid()
        try:
            output = await self.model(prompt=prompt, num_inference_steps=1, guidance_scale=0.0)
            if isinstance(output.images, list) and len(output.images) > 0:
                image = output.images[0]
                buffered = BytesIO()
                image.save(buffered, format="JPEG")
                image_bytes = buffered.getvalue()
                end_time = time.time()
                return image_bytes
            else:
                return None
        except Exception as e:
            return None

async def queue_api_calls(sentence_mapping, character_dict, selected_style):
    """
    Generates images for all provided prompts in parallel using Ray actors.
    Parameters:
        - sentence_mapping (dict): Mapping between paragraph numbers and sentences.
        - character_dict (dict): Dictionary mapping characters to their descriptions.
        - selected_style (str): Selected illustration style.
    Returns:
        dict: A dictionary where keys are paragraph numbers and values are image data in bytes format.
    """
    prompts = []
    for paragraph_number, sentences in sentence_mapping.items():
        combined_sentence = " ".join(sentences)
        prompt = generate_prompt(combined_sentence, sentence_mapping, character_dict, selected_style)
        prompts.append((paragraph_number, prompt))

    num_prompts = len(prompts)
    num_actors = min(num_prompts, 20)
    model_actors = [ModelActor.remote() for _ in range(num_actors)]
    tasks = [model_actors[i % num_actors].generate_image.remote(prompt, f"Prompt {paragraph_number}") for i, (paragraph_number, prompt) in enumerate(prompts)]

    responses = await asyncio.gather(*[ray.get(task) for task in tasks])
    images = {paragraph_number: response for (paragraph_number, _), response in zip(prompts, responses)}
    return images

def process_prompt(sentence_mapping, character_dict, selected_style):
    """
    Processes the provided prompts and generates images.
    Parameters:
        - sentence_mapping (dict): Mapping between paragraph numbers and sentences.
        - character_dict (dict): Dictionary mapping characters to their descriptions.
        - selected_style (str): Selected illustration style.
    Returns:
        dict: A dictionary where keys are paragraph numbers and values are image data in bytes format.
    """
    try:
        loop = asyncio.get_running_loop()
    except RuntimeError:
        loop = asyncio.new_event_loop()
        asyncio.set_event_loop(loop)

    cmpt_return = loop.run_until_complete(queue_api_calls(sentence_mapping, character_dict, selected_style))
    return cmpt_return

gradio_interface = gr.Interface(
    fn=process_prompt,
    inputs=[gr.JSON(label="Sentence Mapping"), gr.JSON(label="Character Dict"), gr.Dropdown(["oil painting", "sketch", "watercolor"], label="Selected Style")],
    outputs="json"
)

if __name__ == "__main__":
    gradio_interface.launch()