Sieve

Detects AI-generated images entirely on-device (WebGPU/WASM). No cloud. No data leaves your browser. Sieve spots AI-generated images while you browse — and does it entirely on your device.

A compact vision model (fine-tuned from Community Forensics, CVPR 2025) runs locally in your browser via WebGPU or WASM. Every image you scroll past gets a quiet confidence badge; images above your threshold (default 65%) are flagged as AI and can be blurred automatically.

WHY SIEVE

• 100% local: no cloud, no APIs, no accounts, no telemetry — images never leave your browser
• Trained for the modern web: GPT Image, Midjourney v7, Flux, Nano Banana and more, tested against heavy compression and thumbnails
• Knows what's NOT AI: screenshots, cartoon frames, game art, YouTube thumbnails, press photos — trained on real-world categories users actually reported
• Honest uncertainty: borderline images get an amber "unsure" badge instead of a false alarm
• One-click reporting: right-click any wrong verdict to pre-fill a public GitHub issue — reports become regression tests and training data for the next model
• Open source (MIT): github.com/Phineas1500/sieve-ai-image-detector

HOW IT WORKS On install, Sieve downloads its detection model once (~45 MB) from a pinned, checksum-verified URL. After that it runs fully offline. Scores are computed with deterministic preprocessing and a numerically self-tested GPU path, so results are consistent across machines and browsers.

Winner of the poidh and 0xSero local-AI detection challenge (36k-image benchmark: 90.9% balanced accuracy clean / 87.1% web-degraded / 85.6% heavily compressed, at a fixed 65% threshold).