app.py

import gradio as gr
import logging
from PIL import Image
import torch
from uroman import Uroman
from transformers import (
    BlipProcessor,
    BlipForConditionalGeneration,
    pipeline,
    AutoTokenizer,
    VitsModel,
)

logging.basicConfig(level=logging.INFO)


processor = BlipProcessor.from_pretrained(
    "Salesforce/blip-image-captioning-large"
)
blip_model = BlipForConditionalGeneration.from_pretrained(
    "Salesforce/blip-image-captioning-large"
).to("cuda" if torch.cuda.is_available() else "cpu")

translation_pipeline = pipeline(
    "translation",
    model="facebook/nllb-200-distilled-600M",
    src_lang="eng_Latn",
    tgt_lang="kor_Hang",
    max_length=200,
    device=0 if torch.cuda.is_available() else -1,
)


tts_ko = VitsModel.from_pretrained("facebook/mms-tts-kor").to(
    "cuda" if torch.cuda.is_available() else "cpu"
)
tok_ko = AutoTokenizer.from_pretrained("facebook/mms-tts-kor")

tts_en = VitsModel.from_pretrained("facebook/mms-tts-eng").to(
    "cuda" if torch.cuda.is_available() else "cpu"
)
tok_en = AutoTokenizer.from_pretrained("facebook/mms-tts-eng")

uroman = Uroman()


def tts(model, tokenizer, text: str):
    roman = uroman.romanize_string(text)
    ids = tokenizer(roman, return_tensors="pt").input_ids.long().to(model.device)
    with torch.no_grad():
        wav = model(input_ids=ids).waveform.squeeze().cpu().numpy()
    return model.config.sampling_rate, wav


def generate(img: Image.Image, lang: str):
    """
    lang == "ko" → 한국어 캡션+음성
    lang == "en" → 영어  캡션+음성
    """
    if img is None:
        raise gr.Error("이미지를 업로드하세요 📷")


    pix = processor(images=img, return_tensors="pt").pixel_values.to(blip_model.device)
    cap_en = processor.batch_decode(
        blip_model.generate(
            pix,
            max_length=128,   # 최대 길이 ↑
            min_length=20,    # 너무 짧게 멈추는 것 방지
            num_beams=5,      # 빔 탐색 품질 ↑ (속도는 조금 느려짐)
            temperature=0.7,  # 다양성 조절
            repetition_penalty=1.1,
        ),
        skip_special_tokens=True
    )[0].strip()


    if lang == "en":
        sr, wav = tts(tts_en, tok_en, cap_en)
        return cap_en, (sr, wav)


    try:
        cap_ko = translation_pipeline(cap_en)[0]["translation_text"].strip()
    except Exception as e:
        logging.error(f"[ERROR] 번역 실패: {e}")
        cap_ko = ""
    if not cap_ko:
        return "번역 오류가 발생했습니다.", None

    sr, wav = tts(tts_ko, tok_ko, cap_ko)
    return cap_ko, (sr, wav)



with gr.Blocks(title="픽보이스(Picvoice)", css="footer {display: none !important;}") as demo:
    gr.Markdown(
        "## 픽보이스(Picvoice)"

    )

    img_state = gr.State()  # 최근 이미지 저장

    input_img = gr.Image(type="pil", label="📷 이미지 업로드")
    caption_box = gr.Textbox(label="📑 텍스트 생성 결과")
    audio_play = gr.Audio(label="🔊 음성 재생", type="numpy")

    with gr.Row():
        ko_btn = gr.Button("한글로 생성🪄")
        en_btn = gr.Button("영어로 생성🪄")


    def store_img(img):
        return img

    input_img.change(store_img, inputs=input_img, outputs=img_state, queue=False)


    ko_btn.click(fn=lambda img: generate(img, "ko"), inputs=img_state, outputs=[caption_box, audio_play])
    en_btn.click(fn=lambda img: generate(img, "en"), inputs=img_state, outputs=[caption_box, audio_play])

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