app.py

import os
import time
import streamlit as st
from PIL import Image
from io import BytesIO
from huggingface_hub import InferenceApi, login
from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer
import torch
from gtts import gTTS
import tempfile

# —––––––– Requirements —–––––––
# This app uses a Hugging Face Transformers version that supports
# the Qwen2.5-Omni architecture via `trust_remote_code`.
# Install using:
# pip install git+https://github.com/huggingface/transformers.git
# and the rest of the requirements listed at the end.

# —––––––– Page Config —–––––––
st.set_page_config(page_title="Magic Story Generator (Qwen2.5)", layout="centered")
st.title("📖✨ Turn Images into Children's Stories (Qwen2.5-Omni-7B)")

# —––––––– Load Clients & Pipelines (cached) —–––––––
@st.cache_resource(show_spinner=False)
def load_clients():
    hf_token = st.secrets["HF_TOKEN"]
    # Authenticate for Hugging Face Hub
    os.environ["HUGGINGFACEHUB_API_TOKEN"] = hf_token
    login(hf_token)

    # 1) BLIP captioning via HTTP API
    caption_client = InferenceApi(
        repo_id="Salesforce/blip-image-captioning-base",
        token=hf_token
    )

    # 2) Load Qwen2.5-Omni model & tokenizer via remote code
    t0 = time.time()
    tokenizer = AutoTokenizer.from_pretrained(
        "Qwen/Qwen2.5-Omni-7B",
        trust_remote_code=True
    )
    model = AutoModelForSeq2SeqLM.from_pretrained(
        "Qwen/Qwen2.5-Omni-7B",
        trust_remote_code=True,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2"
    )
    storyteller = pipeline(
        task="text2text-generation",
        model=model,
        tokenizer=tokenizer,
        device_map="auto",
        temperature=0.7,
        top_p=0.9,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3,
        max_new_tokens=120
    )
    load_time = time.time() - t0
    st.text(f"✅ Story model loaded in {load_time:.1f}s (cached)")
    return caption_client, storyteller

caption_client, storyteller = load_clients()

# —––––––– Helpers —–––––––
def generate_caption(img: Image.Image) -> str:
    buf = BytesIO()
    img.save(buf, format="JPEG")
    resp = caption_client(data=buf.getvalue())
    if isinstance(resp, list) and resp:
        return resp[0].get("generated_text", "").strip()
    return ""


def generate_story(caption: str) -> str:
    prompt = (
        "You are a creative children's-story author.\n"
        f"Image description: “{caption}”\n\n"
        "Write a coherent 50–100 word story that:\n"
        "1. Introduces the main character.\n"
        "2. Shows a simple problem or discovery.\n"
        "3. Has a happy resolution.\n"
        "4. Uses clear language for ages 3–8.\n"
        "5. Keeps each sentence under 20 words.\n"
    )
    t0 = time.time()
    result = storyteller(prompt)
    gen_time = time.time() - t0
    st.text(f"⏱ Generated in {gen_time:.1f}s on GPU/CPU")

    story = result[0]["generated_text"].strip()
    words = story.split()
    if len(words) > 100:
        story = " ".join(words[:100]) + ('.' if not story.endswith('.') else '')
    return story

# —––––––– Main App —–––––––
uploaded = st.file_uploader("Upload an image:", type=["jpg","png","jpeg"])
if uploaded:
    img = Image.open(uploaded).convert("RGB")
    if max(img.size) > 2048:
        img.thumbnail((2048,2048))
    st.image(img, use_container_width=True)

    with st.spinner("🔍 Generating caption..."):
        caption = generate_caption(img)
    if not caption:
        st.error("😢 Couldn't understand this image. Try another one!")
        st.stop()
    st.success(f"**Caption:** {caption}")

    with st.spinner("📝 Writing story..."):
        story = generate_story(caption)

    st.subheader("📚 Your Magical Story")
    st.write(story)

    with st.spinner("🔊 Converting to audio..."):
        try:
            tts = gTTS(text=story, lang="en", slow=False)
            with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as fp:
                tts.save(fp.name)
                st.audio(fp.name, format="audio/mp3")
        except Exception as e:
            st.warning(f"⚠️ TTS failed: {e}")

# Footer
st.markdown("---\n*Made with ❤️ by your friendly story wizard*")