Strip away the branding and every deepfake porn service runs one of four pipelines. This guide explains each stage — the model, the conditioning, the packaging — at the level of detail a researcher, journalist, or policy analyst actually needs, without instructions for reproducing it.
The original 2017-era deepfake ran on a simple, almost crude idea: two autoencoder networks sharing a single encoder. An autoencoder learns to compress a face into a small vector (the encoding) and reconstruct it (the decoding). Train encoder→decoder-A on face A, encoder→decoder-B on face B, then at inference feed A's encodings into B's decoder — out comes face B wearing A's expression and pose. Thousands of source frames, days of GPU training per identity, and output quality that fooled viewers mainly at small sizes and low resolution.
That architecture dominated 2017–2021 because it was the only thing that ran on consumer hardware. Its constraints shaped the genre: it needed many images of the target, it only worked in video (the swap was baked frame-by-frame), and it preserved the body of the original performer — which is why classic deepfake porn is literally face-swapped pornography rather than generated imagery.
The second generation began in August 2022 when Stable Diffusion shipped open weights. Diffusion models inverted the economics: instead of training a network per face, one general model knew every face-like pattern in its training data, and identity was injected at inference time. The technical barrier collapsed from "days of training, ML knowledge" to "download a checkpoint and a 40-megabyte LoRA file."
The undress tool is the highest-volume pipeline and the one responsible for most mass victimization since 2023. Its architecture:
Fig. 1 — The undress pipeline. Note what does NOT change: the face, pose, and background are the victim's actual photograph.
Two properties make this pipeline uniquely harmful. First, input requirements are zero: any clear photo works, no nude source needed, no training. Second, the output inherits credibility — because 70% of the pixels are the victim's real photo, viewers experience it as "a leaked real nude" rather than an obviously synthetic render. Both properties are why inpainting, not face-swap, drives the current wave of school and domestic-abuse cases.
Where inpainting edits a real photo, generation synthesizes an entire scene. The identity layer comes in three flavors:
Because the output is wholly synthetic, it lacks the borrow credulity of an inpainted real photo — generated images carry diffusion artifacts (below). But generation enables what inpainting cannot: scenes, poses, and video content that never existed as photographs of anyone.
The classic deepfake lives on, modernized. The current swap stack:
Output quality now exceeds the old autoencoder pipeline by an order of magnitude while needing a single target photo instead of a video corpus. Remaining weaknesses: profile angles and occlusions (hands passing the face, hair) still produce flicker and identity drift; skin-tone discontinuity at the jaw line remains the most reliable visual tell.
The newest and least mature: video diffusion models (and image-to-video wrappers) generating short explicit clips outright, identity-conditioned like the image pipeline. As of late 2026 output is constrained to seconds-long clips with visible temporal artifacts, but the capability trajectory matches what image diffusion did between 2022 and 2024 — a year behind, closing fast. This pipeline is the one detection researchers watch, because it abolishes the last forensic anchor (real source footage) the other pipelines leave.
| Format | Pipeline inside | User effort | Enforcement surface |
|---|---|---|---|
| Web tools | Inpainting, adapters | Upload + pay | Registrar, host, CDN, payments — high churn |
| Telegram bots | Same backend, chat UI | Send photo to bot | Telegram ToS, crypto payments — resilient |
| Local workflows | All pipelines, incl. video | Install + GPU | Effectively none — no intermediary |
| Commission sellers | Generation + swaps | Send request + crypto | Payment trace only |
The split matters for policy: web tools and bots are disruptable (the domain-churn economics covered in the main reference), while local workflows are not — they run on the user's GPU with open weights, produce the worst material (including CSAM, which no hosted service will touch), and surface only when their outputs are later distributed.
Every pipeline leaves detectable signatures, which is what platform classifiers and forensic analysts key on:
The naive policy answer — run a detector on everything — fails on arithmetic. Classifiers good enough to catch frontier generators produce false-positive rates that would bury moderation queues on real imagery; adversarial post-processing (re-encode, re-photograph, denoise) collapses accuracy further; and the highest-harm material never transits a moderated platform in the first place. Detection is a triage tool inside a broader system whose load-bearing components are provenance standards, payment enforcement, hash-blocking, and the criminal statutes detailed on the law and safety page.