AI-generated pornography has moved from a niche experiment toward a widely available online tool. The images have improved, too. But as the technology gets easier to use, concerns about consent, harassment and moderation are growing alongside it.
More convincing images, more serious consequences
Early AI porn generators often produced bodies that looked physically impossible. Today, some generated images can resemble professional artwork, though the results remain uneven. Better image quality makes the tools more usable, but it also raises the stakes when someone uses them to depict a real person without consent.
Women are the subject of most pornographic deepfakes described in the source, and those images can be used to target people. Twitch creators have continued receiving copies of nonconsensual images through direct messages after the person who made them agreed to delete the originals. A small-town teacher lost her job after parents learned that students had made AI porn in her likeness.
The reported harms also include blackmail and fraud. A 15-year-old boy was targeted with an AI-edited image, while Reddit users were sold explicit pictures of models who did not exist. Workers in adult films and art have also raised concerns about what the technology could mean for their livelihoods.
How Unstable Diffusion built its platform
Unstable Diffusion grew out of online communities using Stable Diffusion, a text-to-image model developed by Stability AI and released as open source. Because only a small part of Stable Diffusion’s training dataset contained adult material, the group recruited volunteers to create datasets for fine-tuning the model.
After bans from Kickstarter and Patreon, the group launched a website with custom image-generating models. It says it raised over $26,000, created a dataset of more than 30 million photographs and now has more than 350,000 users generating over half a million images every day. Those figures are claims made by the group.
Unstable Diffusion describes its mission as supporting uncensored AI art. Its Discord community includes both general-interest images and explicit content. The service is also trying to present itself as a broader art platform, with paid plans, image customization and styles such as digital art, photo and anime.
Quality has improved, but outputs still reflect limits
Testing described in the source found that results were serviceable but unpredictable. Simpler prompts tended to work better, while scenes involving more than two people often produced disturbing results. The models still struggled with the mechanics of sex, sometimes generating unnatural anatomy or impossible positions.
Generated images also showed signs of bias. Prompts for men and women often produced white or Asian people, and prompts for gay porn often defaulted to people of ambiguously LatinX descent with an undercut hairstyle. Body types were narrow by default: men tended to be muscular, and women thin and curvy. The model could produce other body types when prompted, but users had to ask for them explicitly.
These patterns matter because an image model’s defaults shape what users see when they provide little detail. The source links them to possible imbalances in training data, while noting that the exact causes are uncertain. The same issue appeared in a prompt using the word “secretary,” which often produced an Asian woman in a submissive position.
Moderation is central to the business’s future
Unstable Diffusion says it blocks pornographic deepfakes of celebrities and porn depicting characters who appear to be 18 years old or younger. Its system combines blocked words and phrases with an AI model intended to identify and delete prohibited images. The company says its filters are set to be highly sensitive and that it seeks community feedback.
However, the source’s tests found that the filters allowed some celebrity nudes to be generated. After publication, Unstable Diffusion said a bug had allowed content against its policy and that it had fixed the problem and added statistical monitoring. The company also said it had tested the named celebrities again and that they were banned.
The group plans to seek investment, improve its computing capacity and add customization and social-sharing features. Its challenge is to build a mainstream art service while its community and tools remain associated with explicit material. Better image generation can make the product more appealing, but the reported cases show why rules, enforcement and consent are part of the technology’s real-world impact.