DeepNude: the name, the original and what carries it now

The original was a generative adversarial network trained on a narrow dataset, packaged as a Windows and Linux desktop application with a paid tier that removed a watermark. It worked on frontal photographs of women and failed on almost everything else, producing obvious artefacts on any pose it had not seen.

Its author withdrew it days after release, stating that the world was not ready for it. The download links were removed and the official distribution ended. Copies of the binary circulated immediately afterwards and reimplementations followed within weeks.

Technically the original is obsolete. Diffusion models replaced adversarial networks for this task, and every service in our catalogue uses the newer approach. A tool claiming continuity with the 2019 application is claiming something that is not true of its architecture.

In practice it is a web front end over an inpainting pipeline, most often built on publicly available diffusion weights with a segmentation step in front. The engineering effort in these products sits in the interface, the queue and the payment flow rather than in the model.

That is why our scores separate them the way they do. When the generator is effectively common property, what remains to judge is mask control, output resolution, queue behaviour, price structure, retention policy and whether support answers. Those six things are the whole product.

It also explains why the price spread in the catalogue is so wide for such similar output. A service charging several times what its neighbour charges is not running a better model; it is running the same class of model with a different margin.

The term carries a history of non-consensual use, and services that lead with it tend to attract users looking for exactly that. Several of the lowest-scoring entries in our catalogue are also the ones whose marketing leans hardest on the name, and their privacy documentation is correspondingly thin.

Where a service publishes no retention period, no company details and no working support address, a user has no route to have an image removed and no counterparty to address. That combination is what pushes a mark below seven in our ranking more often than output quality does.

The takedown route in Mexico does not depend on the provider cooperating. Material can be reported regardless of where it was generated, and the report page on this site sets out how.

Lawful use only with your own photos, or photos of adults who have given explicit consent. Responsible AI (in Spanish).

• Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models — The paper behind the diffusion inpainting that every tool in this catalogue relies on.
• Kirillov et al., Segment Anything — The segmentation approach used to find a garment region automatically.
• Diario Oficial de la Federación, Ley Olimpia reforms — The Mexican text criminalising dissemination of intimate content, real or simulated, without consent.
• Regulation (EU) 2024/1689, the AI Act — Article 50 sets the disclosure duty for synthetic content that this site follows.
• StopNCII.org — The international route for having non-consensual intimate images removed.
• TAKE IT DOWN Act, S.146 — The United States takedown obligation, relevant where a provider is hosted there.

Technical claims on this page rest on the published research above; legal claims rest on the published texts. Everything else is what the desk observed in its own standardised scenarios, with the date recorded beside each figure.