app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import torch

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

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "vikhyatk/moondream2",
    revision="2025-06-21",
    trust_remote_code=True,
    device_map={"": device}  # or "cpu" / "mps"
)

tokenizer = AutoTokenizer.from_pretrained(
    "vikhyatk/moondream2",
    revision="2025-06-21",
    trust_remote_code=True
)

# Inference function with stream=True and copyable outputs
def caption_image(image: Image.Image, length: str = "normal"):
    # Streaming caption generation
    caption_stream = model.caption(image, length=length, stream=True)["caption"]
    caption = "".join(caption_stream)

    # Token and word counts
    tokens = tokenizer(caption, return_tensors="pt")["input_ids"].shape[1]
    words = len(caption.split())

    return caption.strip(), f"{tokens} tokens", f"{words} words"

# Gradio UI
title = "🖼️ Image Captioner using Moondream 2"
description = (
    "Upload an image and get a generated caption using Moondream 2 (streamed output). "
    "Choose caption length (short, normal, long). Outputs include caption, token count, and word count."
)

demo = gr.Interface(
    fn=caption_image,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Radio(["short", "normal", "long"], value="normal", label="Caption Length")
    ],
    outputs=[
        gr.Textbox(label="Generated Caption", lines=3, show_copy_button=True),
        gr.Textbox(label="Token Count", show_copy_button=True),
        gr.Textbox(label="Word Count", show_copy_button=True)
    ],
    title=title,
    description=description,
    theme="soft"
)

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