AI Slop Detector

AI Slop Detector is a privacy-first browser extension that detects AI-generated images (and sampled video frames) directly on webpages.

All inference runs entirely on your local machine.
No images, frames, or metadata are uploaded or logged.

This project is intended for users who want practical AI-content detection without relying on cloud APIs, telemetry, or third-party services.

---

Overview

Modern AI image generators are increasingly realistic, and many platforms do not clearly label synthetic media. This extension provides a local, model-based signal to help identify likely AI-generated visual content while preserving user privacy.

The system consists of:

• A local FastAPI backend for inference
• A Manifest V3 browser extension that communicates only with localhost

No external network calls are made.

---

Detection Approach

The backend uses an ensemble of two vision classifiers, combining specialized detection with general-purpose coverage.

Models Used (Ensemble)

1. prithivMLmods/OpenSDI-Flux.1-SigLIP2

   • Binary classifier trained specifically to distinguish Flux.1-generated images from real photographs
   • Very effective on Flux-based outputs, including older Grok Imagine variants
   • Robust to moderate post-processing and morphing

2. Ateeqq/ai-vs-human-image-detector

   • General-purpose AI vs human image classifier
   • Trained on a broad mix of recent generators (Midjourney v6+, Stable Diffusion 3.5, GPT-4o images, etc.)
   • Provides strong coverage for non-Flux generators

Ensemble Strategy

• Each model outputs an AI probability score
• The final score is the average of both probabilities
• This reduces single-model bias and improves robustness across generators

Accuracy note
No detector is 100% accurate. Real-world accuracy is typically ~85–95%, depending on generator, image quality, and post-processing. Cutting-edge proprietary models (e.g., current Aurora-based Grok Imagine) may still evade detection.

---

Features

• Automatic scanning of images on any webpage
• Visual overlays:
  • Red border for likely AI-generated content
  • Confidence badge with probability score
• Works on dynamically loaded pages (infinite scroll, lazy loading)
• Basic video support via representative frame analysis
• Adjustable confidence threshold
• No accounts, API keys, or telemetry

---

System Requirements

Backend

• Python 3.10+
• CPU inference supported (GPU optional)

Browser

• Chrome, Edge, or Firefox
• Manifest V3 compatible

---

Installation & Setup

1. Clone the Repository

git clone https://github.com/voidcommit-afk/ai-slop-detector.git
cd ai-slop-detector

2. Install Backend Dependencies

cd backend 
pip install fastapi uvicorn transformers torch pillow

3. Start the Local Backend

uvicorn main:app --reload

Keep this terminal running First launch will download approximately ~800 MB of model weights (one-time)

4. Load the Browser Extension (Chrome)

1. Open chrome://extensions
2. Enable Developer mode
3. Click Load unpacked
4. Select the extension/ directory from the repository

---

5. Test

Visit sites with known AI imagery, such as:

• civitai.com
• lexica.art
• AI art–focused subreddits

Likely AI-generated images should display a red border and a confidence badge.

---

Customization

• Reduce false positives: Increase CONFIDENCE_THRESHOLD in extension/content.js (e.g., to 0.85 or 0.90)
• Detect more subtle AI: Lower the threshold to 0.70 or 0.75
• After changes: Reload the extension in chrome://extensions

Confidence Threshold Examples

• Reduce false positives

CONFIDENCE_THRESHOLD = 0.85 // or 0.90

• Detect more subtle AI artifacts

CONFIDENCE_THRESHOLD = 0.70 // or 0.75

After modifying:

• Reload the extension in chrome://extensions

Privacy & Security

• 100% local inference – All model processing happens on your device
• No outbound network requests from the extension or backend
• No telemetry, analytics, or tracking
• No data persistence or logging
• The browser extension communicates only with localhost:8000

Your images never leave your computer.

---

Known Limitations

• Speed: CPU-only inference can take ~10–30 seconds per batch on average hardware (first run also downloads models)
• Evasion: Very recent or proprietary generators (e.g., current Aurora-based Grok Imagine) may still evade detection
• Video support: Currently samples only a single representative frame
• Probabilistic nature: Output is a confidence score, not definitive proof – false positives and false negatives are possible

Detection accuracy in the wild is typically ~85–95%, varying by generator and post-processing.

---

Contributing

Contributions are welcome! Potential areas include:

• GPU acceleration and improved batching
• Integration of additional ensemble models
• Multi-frame video analysis
• UI/UX enhancements (popup controls, per-site settings)
• Performance optimizations (quantization, caching)

Feel free to open issues or submit pull requests.

---

License

MIT License

---

Built as a learning project with a strong emphasis on privacy, local inference, and realistic expectations of AI detection capabilities.

Enjoy responsibly and keep questioning what's real on the web.