OriginLayer

Multi-Modal Forensic Analysis Platform for Detecting AI-Generated Images

Production-grade forensic analyzer powered by frequency-domain mathematics, vision AI, and 7 novel detection algorithms. Engineered for accuracy.

---

📸 Screenshots

Landing Page

Image Upload & Forensic Scanning

Analysis Results & Confidence Gauge

GAN Fingerprint & Spatial Heatmap

Regional Distribution & Forensic Explanation

---

🔍 What is OriginLayer?

OriginLayer is an AI image forensic platform that determines whether an image is real or AI-generated using a multi-layered detection pipeline. It combines watermark scanning, frequency analysis, vision-language models, and local classifiers into a single unified verdict — with explainable, coordinate-level evidence.

Unlike simple binary classifiers, OriginLayer identifies which AI platform generated the image (DALL-E, Midjourney, Stable Diffusion, Firefly, Gemini) and provides spatial heatmaps showing exactly where artifacts were detected.

---

✨ Key Features

🛡️ Detection Engine Cascade Watermark Detection — SynthID, C2PA, SD watermarks GAN Fingerprint Identification — Platform-level attribution DCT Frequency Analysis — Spectral anomaly detection Gemini Vision LLM — Semantic artifact analysis Local Classifier — Organika/sdxl-detector on CPU	🧠 Novel Research (7 Contributions) FSFN — Frequency-Spatial Fusion Network XAI — Explainable AI with attention maps EUQ — Epistemic Uncertainty Quantification GFI — GAN Fingerprint Identification DACC — Dynamic Accuracy Cross-Calibration SAM — Spatial Artifact Mapping DRWF — Dynamic Reliability Weighted Fusion
🎨 Professional UI Premium dark/light mode with animated toggle Glassmorphism design system Framer Motion micro-animations Responsive across all devices	📊 Forensic Output Coordinate-level artifact detection Interactive spatial heatmaps Confidence gauge with calibration Downloadable PDF forensic reports Analysis history with local storage

---

🏗️ System Architecture

The detection pipeline operates in 4 cascading layers:

Image Upload
    │
    ▼
┌─────────────────────────────┐
│  Layer 1: Watermark Scan    │  SynthID · C2PA · SD Watermark
└──────────────┬──────────────┘
               │
    ┌──────────┼──────────┐
    ▼          ▼          ▼
┌────────┐ ┌────────┐ ┌────────┐
│ Gemini │ │  DCT   │ │ Local  │
│ Vision │ │  Freq  │ │Classif.│
│  LLM   │ │Analysis│ │(SDXL)  │
└────┬───┘ └────┬───┘ └────┬───┘
     └──────────┼──────────┘
                ▼
┌─────────────────────────────┐
│  Multi-Modal Fusion (DRWF)  │  Dynamic Reliability Weighting
└──────────────┬──────────────┘
               ▼
       Forensic Report
  (Verdict · Heatmap · PDF)

---

🛠️ Tech Stack

Layer	Technology
Frontend	Next.js 14 (App Router), React 18, Framer Motion
Styling	CSS Variables, Styled JSX, Glassmorphism Design System
ML Backend	Python FastAPI, PyTorch, HuggingFace Transformers
Detection APIs	Google Gemini Vision, HuggingFace Inference API
Local Models	Organika/sdxl-detector (ViT classifier, 331MB)
Forensics	DCT/FFT Analysis, GAN Fingerprinting, Watermark Detection
Export	html2canvas PDF generation, localStorage history

---

🚀 Quick Start

Prerequisites

• Node.js v20+ — brew install node (macOS) or nodejs.org
• Python 3.9+ — brew install python@3.11 (macOS)

1. Clone & Install

git clone https://github.com/aashish254/ai-image-detection.git
cd ai-image-detection
npm install

2. Configure Environment

cp .env.example .env.local

Edit .env.local with your API keys:

Variable	Required	Get it from
HUGGINGFACE_API_TOKEN	✅	huggingface.co/settings/tokens
GOOGLE_GEMINI_API_KEY	✅	makersuite.google.com
VPT_SERVER_URL	Auto	Default: http://localhost:8100

3. Download ML Models

Note: Model weights (~540MB total) are hosted on HuggingFace Hub — the industry standard for ML model distribution. They are not included in the GitHub repo due to size limits.

Model	HuggingFace Repo	Size	Purpose
SDXL Detector	Organika/sdxl-detector	331MB	Image classifier (ViT)
VPT Artifact Sentinel	aashish254/vpt-artifact-sentinel-v1	208MB	Vision-language forensic model

python3 -m venv .venv
source .venv/bin/activate
pip install huggingface_hub
python scripts/download_model.py

4. Install Python Dependencies

pip install torch torchvision torchaudio transformers peft accelerate \
  pillow fastapi "uvicorn[standard]" python-multipart safetensors \
  sentencepiece huggingface_hub tokenizers psutil

5. Run

Terminal 1 — ML Backend:

source .venv/bin/activate && cd vpt_server && python3 server.py

Terminal 2 — Frontend:

npm run dev

Open http://localhost:3000 🚀

---

💻 Platform Compatibility

Platform	Local Classifier	VPT Model	Notes
macOS (Apple Silicon)	✅ CPU	⚠️ Demo	VPT requires NVIDIA CUDA
Windows + NVIDIA GPU	✅ CUDA	✅ Full	Best experience
Linux + NVIDIA GPU	✅ CUDA	✅ Full	Recommended for production
Any (CPU only)	✅ CPU	⚠️ Demo	Classifier works, VPT in demo mode

---

📁 Project Structure

originlayer/
├── src/
│   ├── app/                          # Next.js App Router
│   │   ├── analyze/page.tsx          # Analysis page
│   │   ├── api/analyze/route.ts      # Detection API endpoint
│   │   ├── features/page.tsx         # Features showcase
│   │   ├── working-model/page.tsx    # How it works
│   │   ├── globals.css               # Theme system & variables
│   │   └── page.tsx                  # Landing page
│   │
│   ├── components/                   # 20+ React components
│   │   ├── AnalysisResults.tsx       # Results dashboard
│   │   ├── ConfidenceGauge.tsx       # SVG confidence arc
│   │   ├── ExplanationPanel.tsx      # Forensic explanation
│   │   ├── SpatialHeatmap.tsx        # Interactive heatmap
│   │   ├── GANIdentificationPanel.tsx# GAN fingerprint display
│   │   ├── Navbar.tsx                # Theme toggle & nav
│   │   └── ...
│   │
│   └── lib/                          # Core detection logic
│       ├── detectors/
│       │   ├── watermark.ts          # SynthID/C2PA/SD detection
│       │   ├── vision-llm.ts         # Gemini Vision integration
│       │   ├── huggingface.ts        # HF API classifier
│       │   └── dct-analysis.ts       # Frequency domain analysis
│       ├── fusion.ts                 # Multi-detector fusion (DRWF)
│       ├── gradcam-analysis.ts       # Forensic report generator
│       ├── gan-fingerprint.ts        # GAN identification
│       ├── frequency-spatial-fusion.ts # FSFN implementation
│       ├── uncertainty.ts            # EUQ implementation
│       └── explainable-ai.ts         # XAI implementation
│
├── vpt_server/                       # Python ML backend
│   ├── server.py                     # FastAPI server
│   └── requirements.txt
│
├── local_models/                     # ML model weights (gitignored)
├── scripts/
│   └── download_model.py             # Model weight downloader
├── docs/                             # Screenshots & diagrams
└── public/                           # Static assets

---

🔬 How Detection Works

1. Watermark Cascade — Scans for embedded digital watermarks (SynthID from Google, C2PA credentials, Stable Diffusion DWT-DCT marks). If found, this provides near-definitive evidence.

2. Vision-Language Analysis — Gemini Vision LLM examines the image for semantic artifacts: unnatural lighting, impossible anatomy, texture inconsistencies, and AI-typical patterns.

3. Frequency Domain Analysis — DCT spectral analysis detects statistical anomalies in frequency coefficients that are invisible to the human eye but characteristic of generative models.

4. Local Classification — The Organika/sdxl-detector ViT model provides a fast binary classification with confidence scoring.

5. Fusion & Calibration — All detector outputs are fused using Dynamic Reliability Weighted Fusion (DRWF), with disagreement-aware confidence calibration producing the final verdict.

---

📄 License

MIT License — see LICENSE for details.

---

_{Built with ❤️ using Next.js, Python, and advanced ML research}