Read this before you rely on it
Who it stops. Who it slows.
Who it cannot stop.
Displayed pixels cannot be made un-copyable — a camera pointed at a screen defeats every purely digital scheme ever shipped, including this one. So the honest question is never “is it unbreakable?” but “what does each attacker actually walk away with, and what evidence do I hold afterwards?” Pick an attacker:
The limits, printed in full
What Archx does not do
- Prevent screenshots. OS-level and extension captures happen above page JavaScript. Archx may degrade them (interlace, readback poisoning), but does not veto them. The 99-case fixed image-transform bench recovered StegoMark in 85 cases and does not include screenshot or camera captures; separate browser tests check viewer capture defenses, not mark recovery from captures. Misses remain possible.
- Stop a camera. Temporal interlace and moiré-unfriendly patterns reduce quality; physics wins eventually. Attribution is the control here.
- Survive AI regeneration. Published diffusion attacks reduce or remove many tested invisible watermark schemes, including neural methods. Archx’s StegoMark has not been measured against full diffusion regeneration; treat it as vulnerable rather than assuming it survives. If your threat is “regenerate my catalog”, you need provenance + legal process, not only marks.
- Prove collusion alone. The assay ranks suspects for investigation; Tardos’ bound puts blind collusion-proof decoding outside a 96-bit channel. Anyone claiming otherwise at this payload size is selling something.
- Replace access control. If everyone may view, everyone may copy at display quality. Archx assumes an authorize decision exists.
Known measured edges
- Adversarial periodic/synthetic content: 0/11 robust recovery in the bench (kept in the corpus deliberately).
- 1-bit re-binarization destroys the mark (a measured dead end for survival at acceptable quality).
- Layout-4 marks under 0.7-scale photo crops can exhaust the shipped bench's 20 s auto-detection budget (pin the layout or raise the budget — pinned searches recover them).
- Robust-profile distortion is ≈28.6 dB PSNR class: visible to pixel-peepers, chosen knowingly (D-12).
maxWidthFractiondefaults to1; an entitled viewer may receive native-width pixels when other caps permit. Set it below1for preview-only delivery; if the fraction would cap below 64 px, delivery is denied rather than rounded up.- Transplanted marked regions carry their pixels’ mark with them — attribution must be corroborated with issuance records, as the guide describes.
When to choose something else instead
| If you need… | Choose | Honest trade-off |
|---|---|---|
| Output-path control (HDCP-class) | EME DRM (Widevine/FairPlay) via a video pipeline | Real output protection for video; licensing/CDM deals, no still-image workflow, heavy integration. Archx can still attribute leaks on top. |
| Provenance & authenticity signalling | C2PA manifest signing | Proves origin/history where verifiers exist; prevents nothing. Complementary to Archx, not a substitute. |
| Highest-fidelity invisible marks at platform scale | Neural watermarking (SynthID-class services) | ~40 dB-class quality; needs ML infrastructure/accounts; equally dead under regeneration. |
| Deep-zoom into gigapixel works | IIIF tiered tile server with resolution-capped access | The established museum pattern; more moving parts than a derivative pipeline. |
| “Nobody may ever see pixels leave” | Don’t publish the image | The only absolute control. Everything else — including this tool — is friction, quality caps, and evidence. |
The engineering rationale behind every one of these boundaries is in the decision atlas — including where our own numbers are unflattering.