Generating from scratch is not editing a photo

The first thing to read is not how much it costs but which action triggers a charge. Ainudez states it most clearly on its pricing page: every generation is counted, including when you are only fixing a face on a result you already have. A correction is therefore not a repair of failed work but a new job at full price.

That rule has a consequence that never appears in the advertising. A diffusion model starts from a random point, so two attempts on the same input give different results, and almost nobody stops at the first attempt. The real cost of an image is the unit price multiplied by the number of tries until you are satisfied — a number only you can calculate, and only afterwards.

The rented engine turns this row upside down. WaveSpeed, the one service that admits it built nothing, is also the one that publishes registered companies with addresses; the services that present the model as their own product are the ones naming nobody. Of the four whose documents open, two write a legal entity and two write only a brand — and it is the brand that the advertising treats as the identity.

That ordering is not about corporate manners. A legal name is the address for a data deletion request, for a payment dispute and for a lawsuit. Without one, all three stop at a mailbox, and a mailbox has no duty to answer. How the six line up on that and every other row is in the comparison of the alternatives that stuck around.

Ainudez shows its interface in several languages, but its four legal documents — terms of service, privacy policy, pricing terms and takedown procedure — are published in English only. The pattern repeats across almost every service of this kind, including ones with localised buttons.

The gap matters precisely because it looks trivial. The translated part is the part that persuades; the untranslated part is the part that binds. For a reader in the Philippines that ordering happens to be harmless — the binding documents are already in a language most people here read — but the habit it reveals is not: what a company translates tells you what it considers worth selling and what it considers worth burying.

Every service whose documents are readable forbids material involving children, and several forbid processing another person's image without permission. How they enforce it is what separates one promise from another. Deep Undress requires a tick-box acknowledging that processing somebody else's image without permission may be illegal, before the job runs. Undresswith describes automatic filters, an age and gender estimation model, human review within 24 hours, and reporting to the US centre for missing children, and requires written consent from anyone appearing in the material.

On the other side stands an admission rarely quoted: the WaveSpeed terms prohibit non-consensual nudity while also stating that the service may produce pornographic content, that the company does not monitor its outputs, and that it is not responsible for them. A written prohibition and actual supervision are two different things, and the difference is carried by the user. What the resulting file turns out to be, and what none of them publish about it, has a page of its own.

This section is almost never read, and on one service its contents collide with the service's own instructions. The UndressHer front page describes a three-step flow whose third step is saving the image to your device. Its terms of service expressly forbid downloading, sharing or distributing generated images, permit use only inside the platform, and warn that a breach can lead to account suspension and legal action.

Anyone following the front-page instructions is automatically breaching the contract they accepted on the way in. This is not just sloppy drafting: when two documents disagree, the one that counts in a dispute is the binding one, not the inviting one. It is also why the comparisons here rest on documents rather than front pages, and why the browser mechanics behind the ones that run in a browser are described from those documents too.

Not one service on this page is a continuation of the 2019 program, and not one holds any right to its name. The name is used because it became a search term: people type it, and a site that puts it in a title gets visits without explaining anything. The full history, from desktop program to generic label, is in the account of what happened in 2019.

The practical consequence is single: a name in a title tells you nothing about who operates the site, which model runs, or where its servers are. Two sites with nearly identical names can belong to two different parties, run different engines and keep files under different rules. The only thing that distinguishes them is the documents they publish themselves — and the same is true of the newer labels, which is why why the names keep changing is worth a page of its own.