AI clothes remover: what these editors actually do

Stage one is segmentation. The model identifies the garment region, either automatically or from a brush stroke the user supplies. Automatic masks are faster and wrong more often, especially where a garment edge sits against a similar tone, or where a hand, a bag strap or a length of hair crosses the boundary. Manual brushing is slower and produces a cleaner cut, which is why every tool that scores well on quality in our catalogue offers it.

Stage two is inpainting. A diffusion model regenerates the masked pixels conditioned on everything outside the mask. It has no information about what the garment covered, so it produces an average of its training distribution constrained by pose, lighting and skin tone at the mask border. This is where the visible failures live: a shoulder that does not line up, a shadow falling the wrong way, a torso lit from a different direction than the face.

Stage three is compositing. The regenerated region is blended back into the original frame. A weak blend leaves a halo, a seam or a sudden change in grain. In our standardised scenarios this is the single most reliable discriminator between tools: at a glance the outputs look similar, and at the mask edge they do not.

None of the three stages recovers anything. There is nothing under the garment in the file to recover. What the reader sees is a generated body attached to a real face, which is precisely why consent and jurisdiction matter more here than the technical quality of any individual tool.

Mask control comes first. A tool that only offers a one-click automatic mask can produce a good result on a simple frontal shot and nothing usable on anything else. Tools with a brush, an eraser and a feather control let a careful user rescue a difficult frame.

Resolution is the second axis, and it is where free tiers are usually crippled. Several services in the catalogue output at a size that hides their own artefacts. A result that looks acceptable at 512 pixels wide falls apart at 1024, and the provider that only offers the smaller size on the free tier is not being generous.

Queue behaviour matters more than raw speed. A tool that returns in twelve seconds when idle and four minutes at peak is harder to work with than one that consistently takes ninety seconds. We record both the observed time and whether the provider publishes any commitment at all, and the absence of a published figure costs points.

Retention is the criterion that moves scores furthest. A provider that states a deletion window in hours, in writing, on a page a reader can find, scores well. A provider whose privacy page says images are kept as long as necessary scores badly, and no amount of output quality compensates for it.

In Mexico the relevant framework is the set of reforms known as Ley Olimpia, which criminalises the dissemination of intimate content without consent and, in the wording adopted across most states, covers material that is real or simulated. Generating a synthetic nude of an identifiable person without their consent falls inside that description in the states that adopted the broader text.

The penalty attaches to the act, not to the tool. No provider in this catalogue offers any protection to a user who uses it on someone else’s photograph, and several state in their own terms that they will cooperate with law enforcement. A tool that does not verify consent is not a tool that shields the person who ignored it.

Where the subject is a minor, the framework is different and far heavier: this is child sexual abuse material regardless of whether any part of the image is synthetic, and there is no jurisdiction in which it is anything other than a serious crime. Every tool in this catalogue prohibits it and several run automated detection.

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.