The Undress AI Index A data desk on the synthetic-intimate-imagery market

Undress AI, measured

Forty domains, three technologies, one legal trendline — the "undress ai" market as a dataset.

Undress AI is a category of web services that synthesize fake nude imagery from a single clothed photograph using diffusion inpainting — generation, not revelation. Measured as a market in November 2026 it counts 40+ churned brands, per-image pricing of $0.10–$5.00, and a legal environment where distribution is criminal in roughly 58 jurisdictions and covered platforms must remove reported content within 48 hours. Everything below is directional, sourced and footnoted — methodology lives on the methodology page.

What the numbers show is a market with a strange inversion: demand is stable, brands are disposable, and law compounds. The operators behind the keyword do not build durable companies — they burn domain names at the rate of one per quarter while reselling the same open-weights pipeline under new labels. Meanwhile the statutes passed between 2022 and 2026 remain in force everywhere they were enacted, which means the dataset's most consequential figures are no longer technical at all: they are removal deadlines, penalty ceilings and hash-registry propagation times.

40+
distinct undress-ai-branded domains observed in WHOIS history since 2023¹
~4 mo
median brand lifespan before domain rotation or suspension²
~90%
of services paywall usable output behind credits or crypto
48 h
statutory removal deadline under the US TAKE IT DOWN Act (2025)
48+
US states & DC criminalizing nonconsensual intimate-image sharing
~96%
of deepfake video online is nonconsensual intimate imagery (2019–24 studies)³

Fig. 1Domain churn: the defining metric

The single most diagnostic statistic in this market is not output quality or pricing — it is registration turnover. Tallying the undress-ai-branded domains appearing in WHOIS records per year shows a market that accelerates even as its brand names vaporize:

0102030 82023 172024 222025 17*2026 YTD
Fig. 1 — Distinct undress-ai-branded domains first observed per year. *2026 through November. Note the 2026 dip coincides with TAKE IT DOWN enforcement, not reduced demand — operators shifted to unbranded churn domains. Source: WHOIS timeline sampling.¹

How the count was made

The chart counts branded registrations only — domains whose names contain the keyword family — first observed in each calendar year, deduplicated where WHOIS fingerprints, shared name servers or cloned page templates link them to a single operator. Two biases pull in opposite directions and are worth stating plainly. Operators recycled some names across registrars, which inflates the raw count slightly; and after mid-2025 a growing share of the market moved to random-string domains that carry no keyword at all, which understates 2026 by an amount we cannot bound tightly. Treat the bars as a floor, not a census.

Why brand names rotate

A four-month median lifespan is not an accident — it is the business model. The forces that kill a branded domain are mostly external and largely predictable:

None of these pressures touches the underlying service. The same render pipeline, credit ledger and Telegram channel simply re-mount under a new label — which is why counting brands overstates the kill rate while counting operators understates the market.

What the 2026 plateau hides

Two readings from the chart matter. First, the market grew through every enforcement wave — demand is inelastic to domain loss because operators treat registrations as disposable. Second, the apparent 2026 plateau is an artifact: after federal removal duties took effect, operators abandoned keyword-branded names for random-string domains, which our sampling method undercounts. The correct interpretation is not that supply shrank; it is that supply went dark to keyword-based measurement.

Fig. 2Market structure, quantified

Table 1 — The undress ai market as measured (November 2026 sample)
MetricValueBasis
Price per image (paid tiers)$0.10 – $5.00Advertised credit pricing across 30+ observed services
Operator GPU cost per image~$0.02 – $0.10Spot-instance inference estimates, SDXL-class inpainting
Free-tier prevalence~70% of servicesOffer at least one free/watermarked render — uploads are the monetization
Payment methods~60% crypto-inclusive; ~40% cardCard acceptance declining as processors enforce policy
Claimed retention ("auto-delete")~85% of servicesAuditable in 0% — no observed service offered proof
Median time-to-paywall1 imageFree tier = one watermarked sample, then credits

The margin column

One row explains the entire market: a render costs an operator single-digit cents of spot GPU time while consumer pricing runs to five dollars — a gross margin between roughly 10× and 250×. That spread, not technical differentiation, is what the branded websites sell. It is also why pricing correlates with nothing: two services running the same open weights differ by an order of magnitude because what is being priced is the customer's willingness, not the computation.

What a free tier actually sells

The ~70% free-tier figure deserves a less charitable reading than the marketing suggests. A zero-price sample is not a subsidy — it is acquisition. What the operator gains from an unpaid render:

The claim nobody audits

The most quoted number in the table is also the least verifiable. Roughly 85% of sampled services advertise automatic deletion of uploads; 0% offer any audit, log transparency or third-party verification mechanism. Given documented archive breaches in this category and the quarterly rotation of its storefronts, the working assumption encoded in every figure on this site is permanent retention — an upload is a disclosure, not a transaction.

Fig. 3The demand side, estimated

What powers that supply is measurable too — approximately. Triangulating keyword-tool volume estimates, autocomplete suggestion counts and referral observations gives a demand picture that is large, stable and mostly informational-criminal in intent:

Table 2 — Demand-side indicators for "undress ai" and adjacent keywords
IndicatorEstimateTrend
Monthly global search volume (keyword family)~1–2MFlat-to-up since 2023; resilient to delisting
Share of searches that are brand-navigational~35–45%Falling as brands churn faster than memory
Share arriving via Telegram/shortlink referrals~25–35%Rising — the churn adaptation
Peak demand periodsNews-cycle spikes (±200–400%)Each enforcement story recruits new demand

Where the traffic enters

The publicity loop

The last table row deserves emphasis: enforcement publicity functions as acquisition marketing. Every takedown story produces a measurable search spike — a Streisand effect that the market's churn model absorbs by design, because a burned brand costs nothing and a news cycle delivers fresh demand for free. It is the structural reason suppression strategies focus on the payment and hosting rails rather than on names — the same dynamic the Electronic Frontier Foundation's deepfakes coverage flags when removal rules are drawn too broadly.

The loop has a second, less visible half. Spikes in demand do not only recruit customers — they recruit victims, because the same coverage that advertises the category also surfaces it to people whose photos are already circulating. Our referral samples show measurable increases in helpline and takedown-tool traffic in the two weeks following major enforcement stories: publicity is simultaneously the market's best marketing channel and its most effective countermeasure.

Fig. 4The enforcement trendline

Against that market data sits a legal dataset moving in the opposite direction. The count of operative legal instruments reached by November 2026:

The offenses that changed the math

Two legal moves rewrote the risk model after 2023. First, creation became a crime in its own right: the UK's 2025 act criminalized producing a synthetic intimate image of an adult without consent, and a minority of US states followed — ending the era when only distributors faced exposure. Second, platforms acquired a clock: the federal US removal duty turned takedown from a courtesy into a 48-hour statutory obligation, the single most consequential number in the dataset for victims.

What the statute count misses

The ~58-jurisdiction figure counts instruments, not outcomes. Statute maxima overstate typical sentences, charging remains rare relative to content volume, and the civil layer — § 6851 actions, state torts, UK and EU claims — runs on timelines of months to years rather than days. Three things the raw count cannot express:

The honest reading: the law now covers the full lifecycle (creation, possession-adjacent conduct, distribution, platform carriage), but enforcement intensity still lags the legislation curve by years. The full legal dataset, with penalty ranges, is tabulated on the legal risk page.

Fig. 5What victims can act on — the useful numbers

For the dataset's most important users, three figures outrank all others: 48 hours (federal platform-removal deadline), 24–72 hours (typical hash-match propagation through StopNCII's partner network), and 1 photo (minimum input modern identity pipelines need to produce a passable composite). What follows is how those numbers combine in practice.

The first seventy-two hours

Victims who act in parallel inside the first two to three days achieve materially better suppression outcomes than sequential filers. The measured order of operations:

  1. Preserve evidence first — URLs, screenshots and timestamps, before any removal request; reports can make source pages disappear.
  2. Hash the image with StopNCII (adults) — the fingerprint is computed on-device, the photo never leaves the phone, and partner platforms then block re-uploads.
  3. File TAKE IT DOWN requests on covered platforms — the 48-hour statutory clock starts at the request.
  4. Request search-engine removal via Google's nonconsensual-imagery removal troubleshooter — delisting starves discovery even where hosting persists.
  5. Report to police and a specialist helpline where the conduct is criminal — the routes below.

Removal and support routes, linked

The verified services behind the numbers: StopNCII (device-side hash-blocking for adults), NCMEC's Take It Down (the same mechanism for images made during minority), the Cyber Civil Rights Initiative (US crisis helpline and attorney referrals), the UK Revenge Porn Helpline, and RAINN for US victim support. National reporting portals include India's cybercrime.gov.in and Italy's Polizia Postale; scam losses — including payments to fraudulent clone services — go to the FTC's ReportFraud portal (what to do if you were scammed). Each is scored by measured speed in the removal table on the legal risk page.

When the subject is or was a minor

Different statute, different route, no tolerance threshold: depictions of minors are CSAM in every jurisdiction in this dataset, including synthetic depictions. Do not retain, forward or examine the material. Report to NCMEC's CyberTipline, and use Take It Down to block re-sharing of images created during minority. The applicable penalty column is measured in decades, not years.

Q&AThe questions readers actually send

Six questions cover almost everything in the inbox; the FAQ page fields the rest, in the same numbered form.

What is undress ai, in one sentence?

A category of services that use diffusion inpainting to synthesize a fake nude body into a clothed photograph of a real person — generation, not revelation: nothing underneath was ever captured. The pipeline is quantified on the methods page.

How many undress ai services exist?

40+ branded domains since 2023 in WHOIS sampling, with a median brand lifespan of about 4 months; unbranded churn domains make the true operator set materially larger. See Fig. 1.

Is undress ai legal?

Distributing nonconsensual synthetic intimate imagery is criminal under US federal law plus ~48 states and DC, the UK, most EU states and Australia. Creating it is now an offense in the UK and a growing minority of US states; depictions of minors are CSAM everywhere. Full table: legal risk.

Do the services really delete uploads?

~85% claim auto-deletion; 0% offered any audit or proof. Treat every upload as permanently retained — the archive is the asset.

How fast can a victim get images removed?

48 hours is the statutory ceiling for covered US platforms under TAKE IT DOWN; StopNCII hash-matching propagates in roughly 24–72 hours. Acting inside the first 72 hours measurably improves suppression.

How much source material does a pipeline need?

Exactly one photograph. Quality degrades only below roughly 512px inputs — standard social-media resolution sits well inside the viable band.

GlossaryTerms used in this briefing

Method note — Domain counts derive from WHOIS timeline sampling of undress-ai-branded registrations; market structure from advertised pricing sampled November 2026; enforcement counts from statute texts and legislative trackers. Figures are directional estimates for a deliberately opaque market — confidence intervals and limitations in the methodology.
Next in the series: the methods page numbers the technical pipeline; legal risk tabulates the statutes; measured quantifies each service archetype; the FAQ answers the most-queried numbers.

¹ Domain counts: WHOIS/RDAP registration-history sampling, deduplicated by operator fingerprint where linkable; unbranded churn domains undercounted by design.

² Median lifespan: first-observed to last-resolved-or-content-abandoned, branded domains only.

³ Share-of-content estimate: synthesis of Deeptrace (2019) and successor studies through 2024; the ratio has remained stable even as absolute volume grew.