Swapping one texture band defeats SynthID detection; tools released as MIT-licensed repo
Kaluga2026 · reddit · 2026-10-12
The author shows how to make a 1536×1024 AI-generated image return "not detected" on a manual SynthID check, with tools and full-size PNGs released in the MIT-licensed synthid-remover repo.
What failed: resizing, geometric tweaks, blur, noise, JPEG recompression, GAN restoration, simulated print-and-camera trips, and SD 1.5 + ControlNet Tile — all still detected.
What worked:
- Generate a donor image with ControlNet Canny (strength 0.9, 30 steps, 640px tiles, overlap 256) — Tile feeds source tones into conditioning so it fails; Canny feeds an edge map, cutting original pixels out of the condition path
- bandmix.py builds a band via two Gaussian blurs (σ=1.2 and 8px) and swaps in the donor's band, keeping 12% of the original band energy (10-12% came back not detected; 15% was detected again)
- Result: mean absolute channel difference 16.91/255, PSNR 20.68 dB; scene recognizable but vegetation gives the swap away
Caveats: verified on one image only, no success rate measured; "not detected" only means that check missed the watermark, not that the image is human-made.
More from Safety
- User says Claude charged his card, demanded government ID, then blocked him — EfficientOrdinary659 · 2026-10-12
- AI safety camps compared: Huang's acceleration vs Kapoor's regulation vs Yudkowsky's alarm — joshua_saxe · 2026-10-12
- UK AI minister says OpenAI-style agent breach of Australia's Medicare could happen in Britain — nordicinst · 2026-10-12
- Raven AI lets cops pinpoint a specific home from a single interior photo, pulling data from Airbnb and Zillow — 404 Media · 2026-10-12
- PDF hidden-text injection is rampant: 795 ICML reviews flagged; check text layer vs render — my-coffee-where · 2026-10-12
- Princeton Launches AI Alignment + Safety Initiative with Inaugural Symposium — HazanPrinceton · 2026-10-12