app.py

import gradio as gr
from huggingface_hub import InferenceClient
from transformers import BlipProcessor, BlipForConditionalGeneration
from PIL import Image
import os
from gtts import gTTS

# 🎨 테마 설정
theme = gr.themes.Soft(primary_hue="blue", secondary_hue="cyan")

# 1. 비전 모델 (로컬)
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
vision_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")

# 2. 텍스트 모델 (안정적인 엔드포인트 설정)
hf_token = os.environ.get("HF_TOKEN")
text_client = InferenceClient(
    model="Qwen/Qwen2.5-7B-Instruct", 
    token=hf_token
)

def describe_photo_like_toeic(image_path):
    if not image_path:
        return "⚠️ Please upload a photo!", "", None
    
    try:
        # 1단계: 사진 분석
        raw_image = Image.open(image_path).convert('RGB')
        inputs = processor(raw_image, return_tensors="pt")
        out = vision_model.generate(**inputs, max_new_tokens=50)
        basic_caption = processor.decode(out[0], skip_special_tokens=True)

        # 2단계: 토익 스피킹 스타일의 초등학생 구어체 프롬프트
        # 사진의 위치(In the middle, on the left 등)를 사용하도록 유도
        prompt = f"""
        Act as an American elementary school student taking a speaking test.
        Describe this image: '{basic_caption}'
        
        Task:
        1. Start with "This is a picture of..."
        2. Describe the main subject in the middle.
        3. Describe what's happening on the left, right, or background.
        4. Use simple kid-friendly words but follow the TOEIC Speaking Part 2 structure.
        5. Finish with a feeling like "Overall, it looks..."
        6. Provide 5 useful expression phrases with Korean meanings.

        Format:
        [Photo Description]
        (Your description here)

        [Useful Phrases]
        1. Phrase - 뜻
        2. Phrase - 뜻
        """

        response = text_client.chat_completion(
            messages=[{"role": "user", "content": prompt}],
            max_tokens=600
        )
        full_text = response.choices[0].message.content

        # 3단계: TTS (사진 묘사 부분만 음성으로)
        audio_path = "description_audio.mp3"
        try:
            description_part = full_text.split("[Photo Description]")[1].split("[Useful Phrases]")[0].strip()
            tts = gTTS(text=description_part, lang='en')
            tts.save(audio_path)
        except:
            tts = gTTS(text="I can describe this photo for you.", lang='en')
            tts.save(audio_path)

        return f"**📷 사진의 첫인상:** {basic_caption}", full_text, audio_path

    except Exception as e:
        return f"❌ Error: {str(e)}", "Please check your API token or connection.", None

# UI 구성
with gr.Blocks(theme=theme) as demo:
    gr.Markdown("# 🎤 초등학생 버전 '토익 스피킹' 사진 묘사하기")
    gr.Markdown("### 사진을 올리면 AI 어린이가 시험을 보듯 조리 있게 설명해 줍니다!")
    
    with gr.Row():
        with gr.Column():
            img = gr.Image(type="filepath", label="🖼️ 묘사할 사진")
            btn = gr.Button("📢 사진 묘사 시작!", variant="primary")
        with gr.Column():
            status = gr.Markdown("사진을 분석하면 영어 설명이 나옵니다.")
            audio = gr.Audio(label="🔊 AI의 설명 듣기", type="filepath", autoplay=True)
            result = gr.Textbox(label="📝 사진 묘사 스크립트", lines=12)

    btn.click(describe_photo_like_toeic, inputs=img, outputs=[status, result, audio])

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