import gradio as gr
from transformers import BlipProcessor, BlipForConditionalGeneration, AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from transformers import AutoTokenizer

# Initialize Blip model for image captioning
model_id = "dblasko/blip-dalle3-img2prompt"
blip_model = BlipForConditionalGeneration.from_pretrained(model_id)
blip_processor = BlipProcessor.from_pretrained(model_id)

# Initialize TTS model from Hugging Face
model_name = "tts-mozilla/tts-ljspeech-multilingual"
tts_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tts_tokenizer = AutoTokenizer.from_pretrained(model_name)
tts = pipeline(task="text2speech", model=tts_model, tokenizer=tts_tokenizer)

def generate_caption(image):
    # Generate caption from image using Blip model
    inputs = blip_processor(images=image, return_tensors="pt")
    pixel_values = inputs.pixel_values
    generated_ids = blip_model.generate(pixel_values=pixel_values, max_length=50)
    generated_caption = blip_processor.batch_decode(generated_ids, skip_special_tokens=True, temperature=0.8, top_k=40, top_p=0.9)[0]

    # Use TTS model to convert generated caption to audio
    audio_output = tts(generated_caption)
    audio_output.save_to_path("generated_audio.mp3")

    return generated_caption, "generated_audio.mp3"

# Create a Gradio interface with an image input, a textbox output, a button, and an audio player
demo = gr.Interface(
    fn=generate_caption, 
    inputs=gr.Image(), 
    outputs=[
        gr.Textbox(label="Generated caption"), 
        gr.Button("Convert to Audio"), 
        gr.Audio(type="player", label="Generated Audio")
    ],
    live=True  # ทำให้ Gradio ทำงานแบบไม่บล็อก
)
demo.launch(share=True)