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

Home › Methods

Technical briefing · November 2026

The pipeline, in numbers

Three stages, one misnomer: what "undress ai" actually computes, and what each stage costs, loses and leaks.

By Henrik Dahl, data editor · Last updated: November 2026

Strip the marketing off an undress ai service and the architecture underneath is unremarkable — a segmentation model, a diffusion inpainter and a blending pass, the same stack any photo editor uses for object removal. What the dataset shows is not novel science but aggressive cost-minimization: operators run commodity open weights at the cheapest viable quality. These are the measured parameters of that pipeline.

3
pipeline stages in every observed implementation
~3.5 B
parameters in a typical SDXL-class inpainting model¹
2–8 s
GPU time per render on datacenter hardware
1
input photograph required by modern pipelines²
~$0.02
marginal operator cost per generated image
50+
diffusion steps sampled per render (typical scheduler settings)

Fig. 1Stage-by-stage, quantified

Table 1 — Pipeline stages with measured characteristics
StageComponentTypical size / costDominant failure mode
1 · SegmentationHuman-parsing + pose models~100–300M params; <1 GPU-secondMask bleed at straps, collars, jewelry
2 · InpaintingFine-tuned diffusion (SD 1.5/SDXL derivatives)~0.9–3.5B params; 2–6 GPU-seconds at 30–50 stepsPlastic texture; missing tattoos, moles, asymmetry
3 · BlendingPoisson/learned composite + face restoreLightweight; <1 GPU-secondBoundary seams; lighting & grain mismatch

The numbers explain the business. A render costs an operator single-digit cents of spot GPU time; credit pricing runs $0.10–$5.00 — a gross margin between roughly 10× and 250× before overhead. That spread, not technical differentiation, is what the branded websites sell.

It also explains the uniformity. With the same open weights underpinning nearly every observed deployment, per-image economics rather than model quality determine what an operator can charge — which is why feature lists across dozens of distinct brands read as near-identical copies of one another.

Fig. 2Generation over time: the quality curve

Quality across engine generations is best expressed as detectability — the share of output surviving expert scrutiny:

0%25%50%75% ~18%pix2pix GAN (2019) ~55%Diffusion inpainting (2023–) ~65%Video pipelines (2025–)
Fig. 2 — Editorial estimate: share of output rated "plausibly real" by trained reviewers across published evaluation corpora. Even the best current generation fails expert review roughly a third of the time — but review happens after distribution.³

Fig. 3Artifact incidence, by region

Table 2 — Where current output fails inspection (incidence across sampled evaluations)
RegionArtifact frequencyTypical defect
Mask boundary~70–80% of samplesSeams, doubled edges, cloth-to-skin transition
Skin detail~60–70%Missing tattoos/moles/scars; uniform "airbrushed" texture
Accessories~40–55%Jewelry/watch absorbed, duplicated or hallucinated
Anatomy~35–50%Asymmetry, merged structures — invented, never accurate to the subject
Provenance~95%+Stripped EXIF, no C2PA credentials

Note the asymmetry between rows: artifacts are frequent enough that forensic detection works, yet the distribution clock outruns review — an image circulates for days before anyone examines it. That gap is the real statistic, and it is why victim-side tooling (hash-matching, statutory takedown) matters more than detector accuracy. Routes and their measured speeds are on the legal risk page.

Fig. 4Input economics

Fig. 5Then vs. now: the 2019→2026 delta

The cleanest way to see what changed is to put DeepNude-era numbers beside the current pipeline:

Table 3 — Category parameters, first generation vs. current
Parameter2019 (GAN era)2026 (diffusion era)Delta
Model parameters~60M (pix2pix)~0.9–3.5B~50×
Render time per image~30–60 s CPU2–8 s GPU~10× faster at scale
Pose coverageFrontal only~90% of typical posesNear-total
Expert-detection escape rate~18%~55–65%~3×
User skill requiredInstall binaryNone (web/bot)Removed entirely
Distribution channelDownload siteApp + bot + API3 parallel channels

Every row moved in the direction of access: more realism, less skill, less time, more channels. Nothing in the 2019→2026 delta favors the defender except provenance — which is why the credential-verification track (C2PA and its successors) matters more than any artifact classifier's benchmark score.

Method note — Parameter counts and compute figures derive from public model cards and benchmark literature; artifact frequencies synthesize published forensic evaluations with our sample observations; monetary figures are spot-instance estimates. Directional, not audited — see methodology.
Continue: the data briefing sizes the market, the legal risk page quantifies enforcement, measured compares archetypes, and the FAQ fields the numbered questions.

¹ SDXL-base ≈ 3.5B parameters; SD-1.5 derivatives ≈ 0.9B — both appear in observed deployments.

² For inpainting (undress ai proper). Face-swap composites need 1–300+ donor photos depending on method.

³ Editorial synthesis of published detector-benchmark corpora and reviewer studies through 2026; not a controlled benchmark of this desk's own.