genai-real-estate-detector

GenAI Detection System for Real Estate Images

Team Beacons | MenaML Winter School 2026 | Track B: Real Estate & Commercial Integrity

Overview

A dual-path detection system for identifying AI-generated real estate images, combining pixel-level forensic analysis with an ensemble of pretrained AI detectors.

Key Features

• Forensic Signal Detection: FFT analysis, texture consistency, color statistics, noise patterns, high-frequency energy analysis
• Ensemble AI Detection: Combines Organika/sdxl-detector and umm-maybe/AI-image-detector
• Tiered Fusion Strategy: Optimized decision logic balancing precision and recall
• Explainable Outputs: Human-readable explanations for each prediction

Performance

Installation

pip install -r requirements.txt

Optional: Install Ollama for VLM support

# macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llava

Usage

python predict.py --input_dir /path/to/images --output_file predictions.json

Arguments

Output Format

{
  "image_name": "example.jpg",
  "authenticity_score": 0.91,
  "manipulation_type": "full_synthesis",
  "vlm_reasoning": "Texture analysis reveals unnaturally uniform surfaces (score: 0.35) typical of AI generation. Color distribution shows limited variation (0.28) inconsistent with natural photography."
}

Score Interpretation

• 0.0 - 0.4: Likely authentic photograph
• 0.4 - 0.6: Borderline / requires manual review
• 0.6 - 1.0: Likely AI-generated or heavily manipulated

Manipulation Types

• authentic: No manipulation detected
• virtual_staging: AI-enhanced or virtually staged
• full_synthesis: Fully AI-generated image

Architecture

Input Image
    │
    ├── Module 1: Forensic Signal Detector
    │   ├── FFT Analysis
    │   ├── Texture Analysis
    │   ├── Color Statistics
    │   ├── Noise Patterns
    │   ├── High-Frequency Energy
    │   └── Local Variance
    │
    ├── Module 2: Ensemble AI Detector
    │   ├── Organika/sdxl-detector (55%)
    │   └── umm-maybe/AI-detector (45%)
    │
    └── Module 3: VLM Reasoner (Ollama LLaVA)
            │
            ▼
    Fusion Module (Tiered Decision Logic)
            │
            ▼
    Output: Score + Type + Explanation

Requirements

• Python 3.8+
• ~2GB disk space for model weights
• CPU inference supported (GPU optional)

License

MIT License

Team

Team Beacons - MenaML Winter School 2026