app.py

import gradio as gr
from transformers import Blip2Processor, Blip2ForConditionalGeneration
from gtts import gTTS
from tempfile import NamedTemporaryFile
from PIL import Image
import torch
import whisper

device = "cuda" if torch.cuda.is_available() else "cpu"

# Load BLIP-2 (smaller model)
processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b").to(device)

# Load Whisper
whisper_model = whisper.load_model("base")

def transcribe(audio_path):
    result = whisper_model.transcribe(audio_path)
    return result["text"]

def ask_image(image, audio):
    question = transcribe(audio)
    inputs = processor(images=image, text=question, return_tensors="pt").to(device)
    generated_ids = model.generate(**inputs)
    answer = processor.decode(generated_ids[0], skip_special_tokens=True)

    tts = gTTS(answer)
    with NamedTemporaryFile(delete=False, suffix=".mp3") as f:
        tts.save(f.name)
        audio_out = f.name

    return answer, audio_out

with gr.Blocks() as demo:
    gr.Markdown("## 🎤🖼️ Ask-the-Image: Ask questions about an image using your voice")

    image_input = gr.Image(type="pil", label="Upload an Image")
    audio_input = gr.Audio(source="microphone", type="filepath", label="Ask a Question (voice)")

    text_output = gr.Textbox(label="Answer")
    audio_output = gr.Audio(label="Answer in Speech")

    btn = gr.Button("Ask")
    btn.click(fn=ask_image, inputs=[image_input, audio_input], outputs=[text_output, audio_output])

demo.launch()