import gradio as gr
from transformers import pipeline
from PIL import Image, ExifTags
import numpy as np
import cv2

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

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

def analyze_image(img_pil):
    img = np.array(img_pil)

    # Blur detection (variance of Laplacian)
    gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
    blur_score = cv2.Laplacian(gray, cv2.CV_64F).var()

    # Noise estimation (std deviation)
    noise_score = np.std(gray)

    # Edge consistency
    edges = cv2.Canny(gray, 100, 200)
    edge_score = np.std(edges)

    # JPEG artifacts (kasar: std dev dari DCT block)
    dct = cv2.dct(np.float32(gray) / 255.0)
    jpeg_artifact = np.std(dct)

    # Metadata kamera
    metadata = "Tidak Ada"
    try:
        exif_data = img_pil._getexif()
        if exif_data:
            metadata = "Ada"
    except:
        pass

    return blur_score, noise_score, edge_score, jpeg_artifact, metadata

def detect_image(img):
    try:
        # Prediksi AI-detector
        result1 = detector(img)
        label1 = result1[0]['label']
        conf1 = round(result1[0]['score'] * 100, 2)

        # Prediksi model general
        result2 = general(img)
        label2 = result2[0]['label']
        conf2 = round(result2[0]['score'] * 100, 2)

        # Analisis teknis
        blur, noise, edge, jpeg_art, metadata = analyze_image(img)

        # --- Voting System ---
        ai_score = 0
        real_score = 0

        # Dari model AI-detector
        if "fake" in label1.lower() or "artificial" in label1.lower():
            ai_score += conf1
        elif "real" in label1.lower() or "human" in label1.lower():
            real_score += conf1

        # Dari metadata
        if metadata == "Ada":
            real_score += 30  # bobot ekstra untuk foto asli
        else:
            ai_score += 20

        # Dari noise & blur
        if noise > 20 and blur > 50:
            real_score += 20
        else:
            ai_score += 10

        # Hasil akhir
        total = ai_score + real_score + 1e-6
        ai_pct = round((ai_score / total) * 100, 2)
        real_pct = round((real_score / total) * 100, 2)

        if ai_pct > real_pct:
            final = f"⚠️ Kemungkinan Besar AI Generated ({ai_pct}%)"
        else:
            final = f"✅ Kemungkinan Besar Foto Asli ({real_pct}%)"

        output = f"""
### Hasil Deteksi:
{final}

**Persentase AI:** {ai_pct}%  
**Persentase Asli:** {real_pct}%

**Model AI-detector:** {label1} ({conf1}%)  
**Model General (ViT):** {label2} ({conf2}%)

**Analisis Kamera & Teknis:**  
- Blur Score: {round(blur,2)}  
- Noise Score: {round(noise,2)}  
- Edge Consistency (STD): {round(edge,2)}  
- JPEG Artifact Level: {round(jpeg_art,2)}  
- Metadata Kamera: {metadata}
"""
        return output
    except Exception as e:
        return f"Terjadi error: {str(e)}"

# UI Gradio
iface = gr.Interface(
    fn=detect_image,
    inputs=gr.Image(type="pil"),
    outputs="markdown",
    title="AI vs Real Image Detector (Hybrid)",
    description="Deteksi AI vs Foto Asli dengan kombinasi model + analisis teknis (blur, noise, edge, metadata)."
)

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