app.py

import streamlit as st
from transformers import pipeline
import torch
from gtts import gTTS
import io

st.set_page_config(page_title="Your Image to Audio Story",
                   page_icon="🦜")
st.header("Turn Your Image📷 to a Short Audio Story🔊 for Children👶")
uploaded_file = st.file_uploader("Select an Image After the Models are Loaded...")

# function part

# Preload models once
@st.cache_resource
def load_models():
    return {
        "img_model": pipeline("image-to-text", "cnmoro/tiny-image-captioning"),
        "story_model": pipeline("text-generation", "Qwen/Qwen2.5-0.5B-Instruct")
    }

models = load_models()


# img2text
def img2text(url):
    text = models["img_model"](url)[0]["generated_text"]
    return text

# text2story
def text2story(text):
    # Define your messages
    prompt = f"Generate a brief 100-word story about: {text}"
    messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
    ]
    response = models["story_model"](
    messages,
    max_new_tokens=100,
    do_sample=True,
    temperature=0.7)[0]["generated_text"]
    story_text = response[2]["content"]
    return story_text

# text2audio
def text2audio(story_text):
    # Create temporary in-memory file
    audio_io = io.BytesIO()
    
    # Generate speech using gTTS
    tts = gTTS(text=story_text, lang='en', slow=False)
    tts.write_to_fp(audio_io)
    audio_io.seek(0)
    
    # Return as dictionary with compatible structure
    return {
        'audio': audio_io,
        'sampling_rate': 16000  # gTTS uses 16kHz by default
    }
    
# Initialize session state variables
if 'processed_data' not in st.session_state:
    st.session_state.processed_data = {
        'scenario': None,
        'story': None,
        'audio': None
    }

if uploaded_file is not None:
    print(uploaded_file)
    bytes_data = uploaded_file.getvalue()
    with open(uploaded_file.name, "wb") as file:
        file.write(bytes_data)
    st.image(uploaded_file, caption="Uploaded Image",
             use_container_width=True)
    # Only process if file is new
    if st.session_state.get('current_file') != uploaded_file.name:
        st.session_state.current_file = uploaded_file.name
        
        # Stage 1: Image to Text
        with st.spinner('Processing image...'):
            st.session_state.processed_data['scenario'] = img2text(uploaded_file.name)
        
        # Stage 2: Text to Story
        with st.spinner('Generating story...'):
            st.session_state.processed_data['story'] = text2story(
                st.session_state.processed_data['scenario']
            )
        
        # Stage 3: Story to Audio
        with st.spinner('Creating audio...'):
            st.session_state.processed_data['audio'] = text2audio(
                st.session_state.processed_data['story']
            )

    # Display results
    # st.image(uploaded_file)
    st.write("Caption:", st.session_state.processed_data['scenario'])
    st.write("Story:", st.session_state.processed_data['story'])

# Keep audio button OUTSIDE file processing block
if st.button("Play Audio of the Story Generated"):
    if st.session_state.processed_data.get('audio'):
        audio_data = st.session_state.processed_data['audio']
        # Convert BytesIO to bytes and specify format
        st.audio(
            audio_data['audio'].getvalue(),
            format="audio/mp3"   # gTTS outputs MP3 by default
        )
    else:
        st.warning("Please generate a story first!")