import gradio as gr
import torch
from diffusers import AutoPipelineForText2Image
from io import BytesIO
from generate_propmts import generate_prompt
from concurrent.futures import ThreadPoolExecutor
import json

# Load the model once outside of the function
model = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo")

def generate_image(text, sentence_mapping, character_dict, selected_style):
    try:
        prompt, _ = generate_prompt(text, sentence_mapping, character_dict, selected_style)
        print(f"Generated prompt: {prompt}")
        output = model(prompt=prompt, num_inference_steps=1, guidance_scale=0.0)
        print(f"Model output: {output}")
        image = output.images[0]
        buffered = BytesIO()
        image_bytes = buffered.getvalue()
        return image_bytes
    except Exception as e:
        print(f"Error generating image: {e}")
        return None

def inference(sentence_mapping, character_dict, selected_style):
    images = {}
    print(f'sentence_mapping: {sentence_mapping}, character_dict: {character_dict}, selected_style: {selected_style}')
    # Here we assume `sentence_mapping` is a dictionary where keys are paragraph numbers and values are lists of sentences
    grouped_sentences = sentence_mapping

    with ThreadPoolExecutor() as executor:
        futures = {}
        for paragraph_number, sentences in grouped_sentences.items():
            combined_sentence = " ".join(sentences)
            futures[paragraph_number] = executor.submit(generate_image, combined_sentence, sentence_mapping, character_dict, selected_style)

        for paragraph_number, future in futures.items():
            images[paragraph_number] = future.result()

    return images

gradio_interface = gr.Interface(
    fn=inference,
    inputs=[
        gr.JSON(label="Sentence Mapping"),
        gr.JSON(label="Character Dict"),
        gr.Dropdown(["Style 1", "Style 2", "Style 3"], label="Selected Style")
    ],
    outputs="json"  
)

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