Media
2 items
Description
Detects AI-generated images on web pages via metadata, C2PA provenance, a SynthID watermark detector, and a visual classifier. SlopGuard flags AI-generated images on the web pages you visit, and runs entirely on your own device. No accounts, no API keys, no servers, nothing uploaded. PLEASE NOTE: this is a large (400M) extension. This is exactly because all of the detection happens locally, so when you install it you have to wait for it to download the detection models. HOW TO USE IT • Click the SlopGuard toolbar icon on any page to scan every eligible image on it (images rendered at roughly 200×200 px or larger). • Or right-click any single image and choose "Check this image for AI" to check just that one. WHAT YOU SEE • A red "AI" or "Probably AI" label, or a yellow "Maybe AI" label, on images judged to be AI-generated — and the image is dimmed to grayscale. • A thin green outline on images that were checked and came back clean. • A grey outline when a check couldn't be completed. HOW IT DECIDES SlopGuard runs several independent checks and stops at the first solid hit: 1. C2PA provenance: reads the cryptographic "content credentials" many generators embed (though are frequently stripped). We don't validate this certificate (i.e. we assume if something says it's AI it's AI). 2. Metadata attribution: EXIF / IPTC / XMP fields that name an AI tool or "AI" as the creator. 3. Byte signatures: fingerprints of known generators (Midjourney, ChatGPT/DALL·E, Adobe Firefly, Stable Diffusion, ComfyUI, and more) and tell-tale generation parameters. This is to catch cases where metadata did exist but got incompletely stripped. 4. SynthID watermark: looks for Google's invisible SynthID watermark directly in the pixels. 5. Visual classifier: a general "does this look AI-generated?" model for images that carry no metadata at all. PRIVATE BY DESIGN Every check runs locally in your browser. The detection models are bundled inside the extension, so nothing is downloaded at scan time and no image, URL, or result is ever sent to the developer or any third party. The only thing stored is a single on/off "Debug mode" setting. SlopGuard acts only when you click. CAVEATS • Visual AI detection is a best guess: expect occasional misses and false positives. • The SynthID check uses an independent community surrogate, not Google's official detector, and detects only Google's invisible watermark. • Images with all metadata completely stripped can only be caught by the visual classifier, if at all. • Only <img> elements are scanned; CSS background images are not. CREDITS On-device detection builds on the work of others: • OpenFake (visual classifier) — Victor Livernoche @ ComplexDataLab. • opensynthid-detect (SynthID surrogate) — fyxme. SlopGuard is not affiliated with, endorsed by, or connected to Google or DeepMind. "SynthID" is a trademark of Google DeepMind, used here only to describe what the watermark check looks for. OPEN SOURCE & LICENSING Non-commercial use only. SlopGuard is free and open source, licensed GPL-3.0-or-later. Source code, build instructions, and the issue tracker are on GitHub: https://github.com/jhpacker/slopguard It bundles third-party models and libraries under their own licenses: • OpenFake (visual classifier) — CC BY-NC 4.0 (NonCommercial) • OpenSynthID (SynthID watermark surrogate) — Apache-2.0 • c2pa-js, onnxruntime-web, exifr — MIT