app.py

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

# 1. Load model AI detector dari HuggingFace
detector = pipeline("image-classification", model="microsoft/resnet-50")

# Fungsi hitung noise (variansi Laplacian)
def estimate_noise(img):
    gray = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2GRAY)
    return cv2.Laplacian(gray, cv2.CV_64F).var()

# Fungsi cek metadata kamera
def has_camera_metadata(img):
    try:
        exif = img._getexif()
        if exif is not None:
            for tag, value in exif.items():
                tag_name = ExifTags.TAGS.get(tag, tag)
                if "Model" in tag_name or "Make" in tag_name:
                    return True
    except:
        pass
    return False

# Fungsi utama deteksi hybrid
def detect_image(img):
    # Resize ke 224x224 untuk model
    img_resized = img.resize((224,224))
    
    # Prediksi AI-detector HuggingFace
    preds = detector(img_resized)
    ai_score = 0
    for p in preds:
        if "artificial" in p['label'].lower() or "fake" in p['label'].lower():
            ai_score += p['score'] * 100
        elif "human" in p['label'].lower() or "real" in p['label'].lower():
            ai_score += (1 - p['score']) * 100
    
    ai_score = max(0, min(100, ai_score))  # jaga range
    real_score = 100 - ai_score

    # Hybrid Adjustment
    # 1. Metadata kamera → tambah bobot real
    if has_camera_metadata(img):
        real_score += 30
        ai_score -= 30

    # 2. Noise → kalau noise tinggi berarti asli
    noise = estimate_noise(img)
    if noise > 500:  # ambang noise
        real_score += 20
        ai_score -= 20

    # Normalisasi agar tetap 0-100
    real_score = max(0, min(100, real_score))
    ai_score = 100 - real_score

    # Final Output
    if real_score == 100:
        label = "🖼️ Gambar ini ASLI 100%"
    elif ai_score == 100:
        label = "🖼️ Gambar ini HASIL AI 100%"
    else:
        label = f"🖼️ Gambar ini {ai_score:.2f}% AI / {real_score:.2f}% Asli"

    return label, f"Noise Score: {noise:.2f} | Metadata Kamera: {'Ada' if has_camera_metadata(img) else 'Tidak'}"

# Gradio UI
demo = gr.Interface(
    fn=detect_image,
    inputs=gr.Image(type="pil"),
    outputs=[gr.Textbox(label="Hasil Deteksi"), gr.Textbox(label="Analisis Teknis")],
    title="Hybrid AI vs Real Image Detector"
)

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