Home › Measured
Comparative dataset · November 2026
Brand names churn; delivery models persist. Four archetypes produce all undress ai output — here are their measured parameters.
We do not compare services by name — in a market where the median brand lives four months, a named comparison is a stale advertisement. What persists, and can be measured, are the four delivery archetypes underneath the branding. Figures below come from our November 2026 sample (30+ observed services plus published vendor and security reporting); confidence notes are in the footnotes and the methodology.
| Dimension | Web apps | Chat bots | Self-hosted pipelines | API wrappers |
|---|---|---|---|---|
| Observed share of market¹ | ~45% | ~30% | ~15% | ~10% |
| Price per image | $0.30–$5.00 | $0.10–$2.00 | ~$0.01–$0.05 (own GPU) | $0.05–$0.50 |
| Photos needed | 1 | 1 | 1 | 1 |
| Card payment accepted | ~55% | ~15% | n/a | ~35% |
| "Auto-delete" claimed | ~90% | ~80% | moot (local) | ~70% |
| Claim auditable | 0% | 0% | n/a | 0% |
| Median lifespan | ~4 mo | ~2–3 mo | stable (code outlives hosts) | ~5 mo |
| Malware association² | Moderate (fake apps/APKs) | Moderate (trojan files) | High (repackaged repos) | Low |
| Legal exposure | Identical across all four — statutes track the act, not the delivery model | |||
Exposure is not uniform — it concentrates differently per archetype. Bars show relative severity on our five measured vectors (higher = worse):
| Vector | Web apps | Chat bots | Self-hosted | API |
|---|---|---|---|---|
| Photo retention | ||||
| Payment trail | ||||
| Malware | ||||
| User identification³ | ||||
| Churn instability |
Two readings from the bars. First, no column trends green: the vector where an archetype scores lowest merely relocates exposure — self-hosted wins on retention and loses on malware. Second, the one uniform row is legal exposure, which sits off-chart at maximum for all four: statutes punish the act and the actor, and no delivery model attenuates it.
For completeness, the response side of the comparison — published detection performance against current-generation synthetic output:
| Option | Published accuracy band⁴ | Coverage | Access |
|---|---|---|---|
| Enterprise detectors (Hive/Sensity-class) | ~90–98% on benchmark corpora | Images + video | Paid; platform-facing |
| Trained human review | ~85–95% given time | All formats | Slow; expert |
| Free public scanners | ~60–80% typical | Stills/links; lags new pipelines | Free |
| Provenance (C2PA) | n/a — verifies authentic, not fake | Growing camera/platform adoption | Free where present |
Two caveats belong on the detector numbers. First, every published band is a moving average measured against the generation that preceded it — accuracy against the newest pipeline always lags until retraining. Second, benchmark corpora skew toward research-friendly samples: compressed, re-shared, watermarked output in the wild degrades every band downward by an amount nobody has cleanly measured. The honest reading of Table 2 is a ceiling, not a guarantee.
¹ Share estimates from branded-domain + referral-observation sampling; self-hosted share is structurally undercounted.
² Association = documented bundling/distribution vectors reported by security vendors, 2023–2026.
³ User identification = composite of payment exposure, platform logs and breach incidence.
⁴ Accuracy figures are corpus-dependent; newer pipelines degrade all bands until detectors retrain.