The origin
DeepNude: how it worked and why it lasted four days
In June 2019 the author pulled the program himself, four days after releasing it. The name kept going without him.
Short answer
DeepNude was one desktop program, released in and withdrawn by its own author a few days later. Since then there has been no new release, no owner's website and nobody maintaining it. If you are looking for "the program" today, the thing you are looking for does not exist — the name now works as a keyword for hundreds of services with no connection to the original project.
This page covers three things in order: what actually existed, how it worked technically, and how a product name turned into the name of a category. The map of what circulates today is in what this guide is about.
A short history
A program that lived four days
The whole story fits into five steps, and all of them happened inside the same week of .
Released as desktop software
It appeared as a site and a download for Windows and Linux. According to the Motherboard report that tested it, it installed and ran like an ordinary Windows application and demanded no technical skill.
Two builds, one watermark
The free build stamped a watermark across the result. The $50 build removed it and put a "FAKE" mark in the upper-left corner instead — a mark the same report described as very easy to crop out or erase in an image editor.
Coverage, then the servers fell
Once it was reported on, traffic ran far past what its author had planned for and the servers went down the same day. Something built for a handful of sales a month was suddenly overrun.
Withdrawn by its author
A day after that coverage, the project's account announced the program was being discontinued. The stated reason was that the probability of misuse was too high, and the closing line was that the world was not yet ready for it.
The copies that stayed
The author said copies would certainly spread but that he did not want to be the one selling them. Shortly afterwards, open-source copies began being removed from public repositories for breaching platform policy.
Two of the sources below are the coverage that triggered the shutdown and the coverage of the shutdown itself: the Motherboard report that first tested the program and the withdrawal notice of 27 June 2019 with the author's statement. The removal of the copies is documented too: repositories hosting reimplementations were disabled for violating the platform's acceptable use policy. All three pieces were published in and can still be read in full.
Behind the word deepfake
What the program was actually doing
It never saw anything behind clothing. It guessed, and every property of its output comes from the way it guessed.
The engine was built on pix2pix, an open-source algorithm based on generative adversarial networks. The model was trained on more than 10,000 photographs and learned to map one image onto another. What came out was not a hidden layer of your source photo but new pixels assembled from thousands of training examples — statistics, not revelation.
From the origin of that dataset came a limit people still misread: the program only worked on photographs of women, and its author explained why — such photographs were easier to collect online. That limit was not an elegant design decision; it was a reflection of the material used to train it.
| What people assume | What actually happened |
|---|---|
| The clothing was taken off | The clothed area was discarded and refilled with the model's guess. The source photo contained no information at all about what was underneath. |
| The result resembled the person | The resemblance came from the face, hair and background, which were left untouched. The replaced region was invented, and it changed on every run. |
| It worked on anyone's photo | The build only functioned on photographs of women, because that is what its training data contained. |
| Everything ran on your own computer | True of the 2019 program. On today's services the computing moved to somebody else's servers, and your file moves there with it. |
That last row is what changed most in seven years. The replacement models are far larger than anything that fits into one installer, so what people use now is almost always a remote interface — and who is running the model at the other end is examined in the piece on which model each one claims to be using.
Today's services use different vocabulary for one family of engines, and the reason two very different operations end up on the same page — generating from text and editing an existing photograph — is explained in why they end up together.
From product to label
How one name became the name of a category
The project died; the search term did not. That is precisely where its value to other people came from.
When the original program disappeared, demand for its name stayed high and no owner had any interest in stopping anyone from using it. Since then the word has attached itself to anything: paid sites, chat bots, installer files, and download pages promising an "old version". What is being sold is not the same product — it is recognition of a single word.
The scale of the category has been counted. Graphika's report recorded 34 synthetic intimate-image providers with more than 24 million unique visitors in a single month in , with referral link spam for such services rising by more than 2,000 per cent on Reddit and X over that year. One dead program left behind one word, and dozens of businesses now use it at once.
Browser services
Paid pages with a tariff, a queue and an account. This is the most common form now, and the comparison of the alternatives that stuck around puts all six on the same rows.
Chat bots
No documents, no named operator, and payments that cannot be pulled back. The rules, where they exist at all, live in a chat bubble either side can delete.
Installer files
Named after the old program and containing anything but — usually a browser window, an activation screen or a second package pulled from elsewhere.
Promises of video
The youngest category and the one that overstates itself most often, because footage is billed frame by frame and the defects only show on playback.
Those four rows have one thing in common: not one descends from the project, and not one is under any obligation to explain its relationship to the name it uses. Sorting the addresses themselves is a job of its own, covered in the domain sort-out.
A word with no owner
Why the name is still being typed seven years on
A product that lasted four days should have taken its name with it. Five things kept the word in circulation instead, and all five are still true.
Nothing was left to own
When a company folds, its name usually stays attached to somebody: a trademark, a liquidator, a buyer. None of that happened here. This was one person's project, it sold for a few days, and it closed without a sale, a handover or a registration anyone has produced since.
So the word entered the language unowned. There is no proprietor to send a cease-and-desist, no register to check a claim against, and no cost to putting it in a page title. That is the first condition for a dead product name turning into a category label, and it is the one condition the original author could not have reversed even if he had wanted to.
The coverage became the asset
Most people who recognise the word never saw the program. What they saw was the reporting about it in , which is also what closed it: the author shut the project down a day after the first article, and the second article recorded him doing it. A story written to warn about the thing is what carried the name to an audience a $50 desktop tool would never have reached.
That is the asset being borrowed now. A service putting the word on its homepage inherits the recognition of the journalism, not the capability of the program — and the two have nothing to do with one another.
What the word still promises
Search behaviour lags seven years behind the market, and people arrive expecting the arrangement: software you download once, a single price rather than a balance, no account, no server, and a visible mark on the result saying it was fabricated. Every one of those four has been reversed. What is on offer now is an account, a metered balance, a remote server that receives your file, and no service among them that commits in writing to marking its output at all.
The gap between those two lists is where most of the money in this category is made, because it is the part nobody has to explain. What actually sits behind the pages still advertising a file to download, and how the original and the copies differ, is taken apart on its own page.
What a page using it is selling
Read a page carrying this name and note what it never states: who operates the service, what one result costs before you register, how long a file stays on a server, and where a complaint goes. Those four are the entire substance of the transaction, and the name answers none of them — it cannot, because it belongs to nobody.
Which turns the word into a filter that works in reverse. The more prominently a page leans on it, the less it has said about itself, and the four questions above stay unanswered no matter how long the copy runs. The money side of what happens after those questions go unasked, and five ways the money goes, is collected separately.
Why the history mitigates nothing
The origin of a technology does not enter the legal analysis. What is weighed is the act: producing sexual material about a person without their consent, and sharing it. The age of the program, who wrote it, and where its servers stood do not alter that. In the Philippines the relevant provisions are Section 12 of the Safe Spaces Act of 2019 on uploading and sharing sexual media without the victim's consent, and the Anti-Photo and Video Voyeurism Act of 2009 on copying, distributing and publishing intimate images of a person taken without consent.
So "I was only curious about the program" stops being a safe position the moment a real person's photograph is involved. Where the person depicted is a child, the Anti-OSAEC and Anti-CSAEM Act of 2022 writes computer-generated images into its definition explicitly, and no argument about how the file was made applies at all.
What exists today
Three browser services using the category
All three run in a browser and have no relationship with the project. What is recorded here is only what they write themselves, including where their own writing is inconsistent.
-
1Pay onceDeep Undress
Sells five access durations as one-off purchases with no auto-renewal, with all sales declared final. On file retention its own documents give three different answers, and the price only appears inside the account.
-
2Open tariffUndresswith
Its price list is complete before an account exists: gem packs from $19.99 to $199.99, paid once, with about 100–110 gems per image by its own page. No document names the company behind it.
-
3Documents closedPornworks
Nothing can be read from outside: requests to the site are answered by a check screen, and even its crawler rules do not open. Its tariff and terms are unreadable, so there is no written promise to save in advance.
The "Open site" buttons are affiliate links: a purchase on the other side earns this desk a commission at no extra cost to you, and it changes nothing here. These three notes come from each service's own documents, not from testing. Where a claim is not written on their side, it is not written here either.
Common questions
About the program and its name
The answers rest on the reporting, on research reports, and on documents anyone can read for themselves.
What exactly was it?
A program for Windows and Linux that replaced the clothed area of a photograph with a model's guess. There was a free build with a watermark and a $50 build that swapped the watermark for a "FAKE" stamp in the corner of the image.
Why was it shut down so fast?
After the coverage, traffic went past what its servers could carry and public attention arrived at the same time. Its author stated that the probability of misuse was too high and stopped sales a day after the report.
So the program cannot be obtained at all now?
There is no official release of it. What remains are uncurated copies with nothing to compare them against, and most files using the name are not even derived from that program.
How is it different from deepfakes in general?
Same family of techniques. A deepfake usually swaps a face in an image or a clip; this program replaced the clothed region of a still photograph. Both produce invented pixels rather than a record of something that happened.
Why did the result change every time it ran?
Because the replaced region is fabricated afresh on each run. There is no body information in the source photo that could be restored, so what appears is a different guess each time.
If the name is free for anyone to use, how do I tell services apart?
Not by the name but by the documents: who operates the service, what it charges before you register, how long it keeps files, and what its refund rule is. On most of them, three of those four questions go unanswered.
Does this desk supply the program?
No. Nothing is hosted here, there are no installation instructions and no images are processed. What exists is a description of what circulates and of what can be checked from outside.