What was “DeepNude”?
A history entry: the 2019 program, its GAN architecture, the shutdown after four days of coverage, the clones that followed, and how the name became a category label.
What was the original program?
“DeepNude” was a Windows and Linux desktop program released in June 2019 that used a conditional GAN, based on the 2016 pix2pix architecture, to draw a synthetic nude over a clothed photo of a woman. Its author withdrew it on June 27, 2019, four days after wide press coverage; the name survived as a generic label for undress apps.
| Item | 2019 original | 2026 category |
|---|---|---|
| Architecture | Conditional GAN (pix2pix, 2016) trained on roughly 10,000 images | Diffusion inpainting on Stable Diffusion-derived checkpoints |
| Delivery | Downloadable desktop binary with a $50 premium tier | Web services, mobile apps, messenger bots, hosted APIs |
| Subjects handled | Women only; failed on men and non-frontal poses | Any adult body type; minors blocked by classifiers in reviewed tools |
| Legal frame | No specific US statute; general NCII law | TAKE IT DOWN Act (2025), 15 U.S.C. §6851, state nudification bans |
How did the 2019 program work?
The program ran three stages. A first network drew a rough mask over the clothing, a second inferred an anatomical sketch, and a third pix2pix generator rendered skin texture into the sketch.
Each stage was a conditional GAN: a generator mapped an input image to an output image while a discriminator judged whether the pair looked real. Because the training set was small and one-sided, results were blurry, frontal-only and full of artifacts.
We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems.
— Isola et al., pix2pix paper (arXiv:1611.07004, 2016), accessed September 3, 2026
Why was it shut down?
Coverage on June 23, 2019 drove so many downloads that the servers failed. On June 27 the author posted that the risk of misuse was too high and that “the world is not yet ready” for the software, closed sales and promised refunds. Copies had already spread; within weeks the code was reverse-engineered and uploaded to public repositories.
What replaced it?
Between 2019 and 2022, web clones with the same GAN backbone ran on the leaked weights. After Stable Diffusion shipped in August 2022, the category moved to diffusion inpainting, which handles any pose and body type with far fewer artifacts.
The reviewed tools in Undress AI rankings all use diffusion pipelines; none descends from the 2019 code. Many services that copied the name have since closed, as documented in undress apps that shut down.
Why does the name still matter?
The word became shorthand in press, hearings and complaints for any tool of this kind, so people search for it when they mean the category. This glossary defines it to state what the program was, that it no longer exists as a product, and that any current service using the name is unrelated to the 2019 author.
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
- Isola et al., pix2pix (2016) — arXiv:1611.07004 via Hugging Face, accessed September 3, 2026.
- Goodfellow et al., “Generative Adversarial Networks” (2014) — arXiv:1406.2661 via Hugging Face, accessed September 3, 2026.
- CompVis/stable-diffusion repository — GitHub, accessed September 3, 2026.
- TAKE IT DOWN Act, Public Law 119-12 — GovInfo, accessed September 3, 2026.
- 15 U.S.C. §6851 — Cornell LII, accessed September 3, 2026.
- 18 U.S.C. §2256 — Cornell LII, accessed September 3, 2026.
FAQ
Does the original 2019 program still exist?
Not as a product. Sales stopped on June 27, 2019, and the author has published nothing since. Copies and reverse-engineered repositories circulated afterwards, but they run an outdated GAN that fails on most photos. Websites using the name in 2026 are unrelated projects that borrow it as a keyword.
What technology did it use?
A three-stage pipeline of conditional GANs based on the 2016 pix2pix paper: one network masked the clothing, one produced an anatomical sketch and one rendered texture. It was trained on roughly 10,000 images of women, which explains why it failed on men, side views and unusual clothing.
Why was it withdrawn after four days?
Press coverage on June 23, 2019 caused a download surge that crashed the servers and made the misuse potential obvious. On June 27 the author wrote that the probability of abuse was too high, ended sales and offered refunds. The decision could not stop copies already downloaded.
How do today’s undress apps differ from it?
They use diffusion inpainting on Stable Diffusion-derived checkpoints instead of a GAN, handle any pose or body type, run in a browser or messenger rather than as a desktop binary, and the reviewed tools add consent checks and minor-detection classifiers. Output quality is far higher; so is the legal risk.
Is downloading an old copy legal?
Possessing the binary is not itself a crime in the United States, but using it on an identifiable person without consent can violate the TAKE IT DOWN Act and 15 U.S.C. §6851, and any minor in the image triggers 18 U.S.C. §2256. Old copies also carry malware risk.
Why does a review site define this term?
Because readers search for the name when they mean the category. The entry states what the program was, that it ended in 2019 and that current tools share only the idea. Reviews on this site cover live third-party services under one consent-based protocol; the site runs no generator.