app.py

import torch
from transformers import (
    pipeline,
    BlipProcessor,
    BlipForConditionalGeneration,
)
import gradio as gr

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

# -------- VQA PIPELINE --------
vqa = pipeline(
    "visual-question-answering",
    model="Salesforce/blip-vqa-base",
    device=0 if DEVICE == "cuda" else -1,
)

# -------- CAPTION MODEL (Manual, more stable) --------
caption_processor = BlipProcessor.from_pretrained(
    "Salesforce/blip-image-captioning-base"
)
caption_model = BlipForConditionalGeneration.from_pretrained(
    "Salesforce/blip-image-captioning-base"
).to(DEVICE)

def generate_caption(image):
    if image is None:
        return ""

    inputs = caption_processor(image, return_tensors="pt").to(DEVICE)
    out = caption_model.generate(**inputs)
    caption = caption_processor.decode(out[0], skip_special_tokens=True)
    return caption

def answer_question(image, question):
    if image is None:
        return "Please upload an image first."
    if not question or not question.strip():
        return "Please type a question about the image."

    result = vqa(question=question, image=image)
    return result[0]["answer"]

with gr.Blocks() as demo:
    gr.Markdown("# BLIP Captioning + Visual Question Answering")
    gr.Markdown(
        "Upload an image to generate a caption, then ask a question about the image for an answer."
    )

    with gr.Row():
        image_in = gr.Image(type="pil", label="Upload an image")
        with gr.Column():
            caption_out = gr.Textbox(label="Caption (auto-generated)", lines=2)
            answer_out = gr.Textbox(label="Answer", lines=2)

    question_in = gr.Textbox(
        label="Question",
        placeholder="e.g., What is in the image? How many people are there?",
    )

    with gr.Row():
        clear_btn = gr.Button("Clear")
        answer_btn = gr.Button("Submit")

    image_in.change(fn=generate_caption, inputs=image_in, outputs=caption_out)
    answer_btn.click(fn=answer_question, inputs=[image_in, question_in], outputs=answer_out)
    clear_btn.click(fn=lambda: (None, "", "", ""), outputs=[image_in, question_in, caption_out, answer_out])

    gr.Markdown("**Note:** Outputs may be incorrect. Do not use for medical/legal decisions.")

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