app.py

import streamlit as st
from transformers import pipeline
from PIL import Image
import io, textwrap, numpy as np, soundfile as sf

# ------------------ Streamlit Page Configuration ------------------
st.set_page_config(
    page_title="Picture to Story Magic",  # App title on browser tab
    page_icon="🦄",                       # Fun unicorn icon
    layout="centered"
)

# ------------------ Custom CSS for a Colorful Background ------------------
st.markdown(
    """
    <style>
    body {
        background-color: #FDEBD0;  /* A soft pastel color */
    }
    </style>
    """,
    unsafe_allow_html=True
)

# ------------------ Playful Header for Young Users ------------------
st.markdown(
    """
    <h1 style='text-align: center; color: #ff66cc;'>Picture to Story Magic!</h1>
    <p style='text-align: center; font-size: 24px;'>
      Hi little artist! Upload your picture and let us create a fun story just for you! 🎉
    </p>
    """,
    unsafe_allow_html=True
)

# ------------------ Lazy Model Loading ------------------
def load_models():
    """
    Lazy-load the required pipelines and store them in session state.
    Pipelines:
      1. Captioner: Generates descriptive text from an image using a lighter model.
      2. Storyer: Generates a humorous children's story using aspis/gpt2-genre-story-generation.
      3. TTS: Converts text into audio.
    """
    if "captioner" not in st.session_state:
        # Use the "base" version for faster/cost-effective captioning.
        st.session_state.captioner = pipeline(
            "image-to-text",
            model="Salesforce/blip-image-captioning-base"
        )
    if "storyer" not in st.session_state:
        st.session_state.storyer = pipeline(
            "text-generation",
            model="aspis/gpt2-genre-story-generation"
        )
    if "tts" not in st.session_state:
        st.session_state.tts = pipeline(
            "text-to-speech",
            model="facebook/mms-tts-eng"
        )

# ------------------ Caching Functions ------------------
@st.cache_data(show_spinner=False)
def get_caption(image_bytes):
    """
    Convert the image bytes into a smaller image to speed up captioning,
    then return the generated caption.
    """
    image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
    # Resize the image (preserving aspect ratio) to only 256x256 for faster processing.
    image.thumbnail((256, 256))
    caption = st.session_state.captioner(image)[0]["generated_text"]
    return caption

@st.cache_data(show_spinner=False)
def get_story(caption):
    """
    Generate a humorous and engaging children's story using the caption.
    The prompt instructs the model to produce a playful story (50-100 words).
    We lower max_new_tokens to 80 so that it generates its text faster.
    """
    prompt = (
        f"Write a funny, warm, and imaginative children's story for ages 3-10, 50-100 words, "
        f"in third-person narrative, as if the author is playfully describing the scene in the image: {caption}. "
        "Explicitly mention the exact venue or location (such as a park, school, or home), describe specific characters "
        "(for example, a little girl named Lily or a boy named Jack), and detail the humorous actions they perform. "
        "Ensure the story is playful, engaging, and ends with a complete sentence."
    )
    raw_story = st.session_state.storyer(
        prompt,
        max_new_tokens=80,   # Reduced token generation for faster response
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
        return_full_text=False
    )[0]["generated_text"].strip()
    words = raw_story.split()
    return " ".join(words[:100])

@st.cache_data(show_spinner=False)
def get_audio(story):
    """
    Convert the generated story text into audio.
    The text is split into 300-character chunks to reduce repeated TTS calls,
    the audio chunks are concatenated, and then stored in an in-memory WAV buffer.
    """
    chunks = textwrap.wrap(story, width=300)
    audio_chunks = [st.session_state.tts(chunk)["audio"].squeeze() for chunk in chunks]
    audio = np.concatenate(audio_chunks)
    buffer = io.BytesIO()
    sf.write(buffer, audio, st.session_state.tts.model.config.sampling_rate, format="WAV")
    buffer.seek(0)
    return buffer

# ------------------ Main App Logic ------------------
uploaded_file = st.file_uploader("Choose a Picture...", type=["jpg", "jpeg", "png"])
if uploaded_file is not None:
    try:
        load_models()  # Ensure models are loaded once
        image_bytes = uploaded_file.getvalue()
        # Display the user-uploaded image
        image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
        st.image(image, caption="Your Amazing Picture!", use_column_width=True)
        st.markdown("<h3 style='text-align: center;'>Ready for your story?</h3>", unsafe_allow_html=True)
        
        if st.button("Story, Please!"):
            with st.spinner("Generating caption..."):
                caption = get_caption(image_bytes)
            st.markdown("<h3 style='text-align: center;'>Caption:</h3>", unsafe_allow_html=True)
            st.write(caption)
            
            with st.spinner("Generating story..."):
                story = get_story(caption)
            st.markdown("<h3 style='text-align: center;'>Your Story:</h3>", unsafe_allow_html=True)
            st.write(story)
            
            with st.spinner("Generating audio..."):
                audio_buffer = get_audio(story)
            st.audio(audio_buffer, format="audio/wav", start_time=0)
            st.markdown(
                "<p style='text-align: center; font-weight: bold;'>Enjoy your magical story! 🎶</p>",
                unsafe_allow_html=True
            )
    except Exception as e:
        st.error("Oops! Something went wrong. Please try a different picture or check the file format!")
        st.error(f"Error details: {e}")