AKAnswered by Anna Kowalski · Last updated: November 2026
Short answer
AI nude imagery is made by diffusion models — neural networks that generate images by gradually removing noise. For real-person nudes, three pipelines dominate: inpainting (mask the clothing in a real photo, regenerate that region), face swapping (paste a real face onto existing footage), and identity-conditioned generation (teach the model a face, then generate any scene). The first needs only one ordinary photo, which is why it's the one behind the mass-harm cases.
What is a diffusion model, in plain terms?
Picture a training process where real photographs are slowly corrupted with static until they're pure noise — while a network learns, at every corruption level, what the slightly-cleaner version looked like. Run that learning in reverse, starting from pure static plus an instruction, and the network hallucinates a brand-new image that looks like it came from its training data. That's diffusion. It's the engine inside Stable Diffusion, SDXL, and every image generator you've heard of — and the one inside every AI nude tool.
The key consequence: the model doesn't retrieve a body, it predicts one. An AI nude of a real person contains zero pixels of that person's actual body — the anatomy is a statistical guess conditioned on the visible evidence. It can be plausible and completely invented at the same time.
How does inpainting — the "undress" pipeline — work?
The highest-volume pipeline and the one behind the 2023–2026 mass-victimization wave. Four stages:
Segmentation. A model trained specifically on clothed bodies draws a mask over the garment region — tracking limbs, preserving face, hair, jewelry, background.
Inpainted diffusion. A checkpoint fine-tuned on nude datasets regenerates only the masked pixels, conditioned on everything around them: pose, build, skin tone, lighting, camera grain.
Blending. Edge feathering and color-matching unify the synthesized region with the untouched photo.
Delivery. Output to user; the input photo — and usually the output — retained by the operator.
The undress pipeline: most of the output is the victim's own photograph — which is what makes it convincing.
That last row is the crucial property. Because the model only touches the masked region, everything else — the person's real face, their actual bedroom, their real posture — passes through untouched. The output inherits the credibility of a genuine photograph, which is precisely why it lands as "a leaked nude" in the mind of anyone who sees it.
How do face swaps and conditioned generation differ?
The other two pipelines for real-person imagery:
Face swapping re-faces existing photos or video frames — a single-model swap preserves the original's expression, lighting, and motion while substituting the target's identity. It's the classic "deepfake" technique, now one forward pass per image rather than days of per-person training. Its tell: identity drift under occlusion and profile angles.
Identity-conditioned generation teaches a general model a specific face — either via a small fine-tune file (LoRA, trained on 10–40 photos) or a face-reference adapter that needs one photo and no training. The model then generates any invented scene wearing that face. This is the "commission" pipeline for made-to-order material.
Plus a third category entirely: unconditioned generation — pure text-to-image synthesis of invented people, no real likeness involved. Same engine, no victim, entirely different legal and ethical position — the distinction the main answer draws in its three-row table.
Why does the output look real?
Four reasons, in order of importance:
Inherited context. In inpainting, 70%+ of the final image is a real photograph — real lighting, real grain, real room. The human eye reads the whole frame as authentic.
Learned anatomy. The model fills from a huge corpus of what bodies look like — plausible proportions, consistent skin tone, no obvious render tells at thumbnail size.
Trained priors. Diffusion models are very good at "what plausibly goes here" — that is literally what the training objective is.
Viewers' expectations. A fake that looks like a phone photo faces less scrutiny than a fake that looks like a render — and these tools specifically output the phone-photo look because they preserve the original's artifacts.
Why does quality vary so much between outputs?
Three variables control it: input quality (sharp, front-facing, well-lit photos convert dramatically better than candids), mask quality (a contour-following mask blends seamlessly; a blocky one smears limbs — the signature of cheap services), and checkpoint quality (fine-tunes trained on better corpora produce less of the "airbrushed mannequin" look that both human reviewers and forensic detectors key on).
The tell that survives: generated regions trend symmetric, smooth, and average — the model's learned mean. Under scrutiny the "too clean" quality persists even in good output, and under forensic analysis the mask boundary and generator fingerprints are legible regardless. More on detection in the FAQ.