import gradio as gr
from transformers import pipeline
from PIL import Image, ExifTags, ImageStat, ImageFilter
import numpy as np

# Model utama untuk deteksi AI
detector = pipeline("image-classification", model="umm-maybe/AI-image-detector")

# Model tambahan general classifier (backup)
general = pipeline("image-classification", model="google/vit-base-patch16-224")

def analyze_noise(img):
    gray = img.convert("L")
    arr = np.array(gray)
    return np.std(arr)  # Standar deviasi → noise

def analyze_blur(img):
    gray = img.convert("L")
    arr = np.array(gray)
    lap = cv2.Laplacian(arr, cv2.CV_64F).var()
    return lap  # Variansi Laplacian → blur

def metadata_score(img):
    try:
        exif = img._getexif()
        if exif is None:
            return 1  # Tidak ada metadata → kemungkinan AI
        for tag, value in exif.items():
            decoded = ExifTags.TAGS.get(tag, tag)
            if decoded == "Make" or decoded == "Model":
                return 0  # Ada kamera → kemungkinan asli
        return 1
    except:
        return 1

def detect_image(img: Image.Image):
    try:
        # Prediksi AI-detector
        result1 = detector(img)
        label1 = result1[0]['label']
        conf1 = result1[0]['score']  # 0-1

        # Prediksi general model
        result2 = general(img)
        label2 = result2[0]['label']
        conf2 = result2[0]['score']

        # Analisis blur & noise
        noise = analyze_noise(img)
        try:
            import cv2
            blur = analyze_blur(img)
        except:
            blur = 0

        # Metadata
        meta = metadata_score(img)

        # Weighted skor akhir (lebih sensitif terhadap AI photorealistic)
        ai_score = conf1 * 0.6 + meta * 0.2 + (1 - min(noise/100,1)) * 0.1 + (1 - min(blur/1000,1)) * 0.1
        ai_score = min(max(ai_score, 0), 1)  # Clamp 0-1
        human_score = 1 - ai_score

        ai_percent = round(ai_score * 100, 2)
        human_percent = round(human_score * 100, 2)

        output = f"""
### Hasil Deteksi:
🖼️ Gambar ini {ai_percent}% AI / {human_percent}% Asli

**Model AI-detector:** {label1} ({round(conf1*100,2)}%)  
**Model General (ViT):** {label2} ({round(conf2*100,2)}%)  
**Blur Score:** {round(blur,2)}  
**Noise Score:** {round(noise,2)}  
**Metadata Kamera:** {"Ada" if meta==0 else "Tidak Ada"}
"""
        return output
    except Exception as e:
        return f"Terjadi error: {str(e)}"

iface = gr.Interface(
    fn=detect_image,
    inputs=gr.Image(type="pil"),
    outputs="markdown",
    title="Hybrid AI vs Foto Asli Detector",
    description="Upload foto untuk mendeteksi persentase AI vs foto asli. Lebih sensitif terhadap AI photorealistic."
)

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