Home › Methods
Technical briefing · November 2026
Three stages, one misnomer: what "undress ai" actually computes, and what each stage costs, loses and leaks.
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.
| Stage | Component | Typical size / cost | Dominant failure mode |
|---|---|---|---|
| 1 · Segmentation | Human-parsing + pose models | ~100–300M params; <1 GPU-second | Mask bleed at straps, collars, jewelry |
| 2 · Inpainting | Fine-tuned diffusion (SD 1.5/SDXL derivatives) | ~0.9–3.5B params; 2–6 GPU-seconds at 30–50 steps | Plastic texture; missing tattoos, moles, asymmetry |
| 3 · Blending | Poisson/learned composite + face restore | Lightweight; <1 GPU-second | Boundary 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.
Quality across engine generations is best expressed as detectability — the share of output surviving expert scrutiny:
| Region | Artifact frequency | Typical defect |
|---|---|---|
| Mask boundary | ~70–80% of samples | Seams, 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.
The cleanest way to see what changed is to put DeepNude-era numbers beside the current pipeline:
| Parameter | 2019 (GAN era) | 2026 (diffusion era) | Delta |
|---|---|---|---|
| Model parameters | ~60M (pix2pix) | ~0.9–3.5B | ~50× |
| Render time per image | ~30–60 s CPU | 2–8 s GPU | ~10× faster at scale |
| Pose coverage | Frontal only | ~90% of typical poses | Near-total |
| Expert-detection escape rate | ~18% | ~55–65% | ~3× |
| User skill required | Install binary | None (web/bot) | Removed entirely |
| Distribution channel | Download site | App + bot + API | 3 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.
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.