What is inpainting?
The operation at the center of every undress app: how a diffusion model fills a mask, which parameters control the result, and what it cannot do.
What does inpainting do?
Inpainting is the technique of filling a masked region of an image with new pixels that match the surrounding content. Classical methods copied nearby texture; since 2021 diffusion models generate the fill from a text prompt, and Stability AI’s 2022 inpainting checkpoint trained for 200,000 extra steps at 512×512 made the method the standard engine of undress apps.
| Term | What it does | Where used | Limits |
|---|---|---|---|
| Inpainting | Regenerates only the masked region | Object removal, retouching, undress apps | Fill is invented; seams and anatomy errors |
| Outpainting | Extends the canvas beyond the original borders | Aspect-ratio changes, backgrounds | Same invention problem outside the frame |
| img2img | Re-renders the whole image from a noised copy | Style transfer, variations | Changes everything unless a mask is added |
| Face swap | Replaces facial identity with a reference face | Video and photo identity edits | Different pipeline; face only |
How does diffusion inpainting work?
A diffusion model learns to remove noise from images step by step. For inpainting, the pipeline takes three inputs: the image, a binary mask and a prompt. It encodes the image into latent space, adds noise, and at each denoising step keeps the unmasked latents fixed while regenerating the masked ones under the prompt.
In the Diffusers pipeline the defaults are 50 steps, guidance 7.5 and strength 1.0, meaning the masked area is fully regenerated; the base model and any LoRA adapters decide what the fill looks like.
By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond.
— Rombach et al., latent diffusion paper (arXiv:2112.10752, 2021), accessed September 3, 2026
Why does every undress app depend on it?
An undress app is an inpainting pipeline with an automatic mask. A segmentation model marks the clothing, the mask goes to an inpainting checkpoint fine-tuned on nude imagery, and a fixed internal prompt requests skin. Nothing is recovered from under the fabric; the region is invented.
Test notes in the clothes-removal category show the typical failures: mismatched skin tone, duplicated navels, merged hands. The how undress apps work article illustrates each step.
Which parameters change the result?
Mask quality matters most: a mask that misses a sleeve leaves fabric; one that covers the hands makes the model redraw them. Strength below 1.0 keeps some original pixels, useful for retouching but not for full replacement.
Guidance scale between 5 and 9 trades prompt adherence for naturalness. Resolution is fixed by the checkpoint: 512×512 for Stable Diffusion 1.x and 2.x, 1024×1024 for SDXL.
What can inpainting not do?
It cannot reveal hidden information, restore a face it never saw or reproduce a real body from a clothed photo. It also cannot check consent: the mask and prompt are indifferent to who is in the picture, which is why consent controls and NSFW and age classifiers sit outside the model in the tools we review.
Inpainting a real person into an intimate image without consent creates a digital forgery under the TAKE IT DOWN Act.
Consent Note
This site is a publisher: it reviews third-party tools and does not create, edit, accept or store images. Docs entries are general information, not legal advice.
Sources
- Rombach et al., latent diffusion (2021) — arXiv:2112.10752 via Hugging Face, accessed September 3, 2026.
- Model card: stabilityai/stable-diffusion-2-inpainting — Hugging Face, accessed September 3, 2026.
- Diffusers docs: inpainting pipeline — Hugging Face, accessed September 3, 2026.
- facebookresearch/segment-anything — GitHub, accessed September 3, 2026.
- TAKE IT DOWN Act, Public Law 119-12 — GovInfo, accessed September 3, 2026.
- How we rate — Clothoff AI.
FAQ
Is inpainting the same as Photoshop’s content-aware fill?
Same goal, different method. Content-aware fill copies and blends texture that already exists in the image, so it works for skies and walls. Diffusion inpainting generates new content from a prompt and training data, so it can draw objects or bodies that were never in the picture.
Does inpainting reveal what is under clothing?
No. The masked region is discarded and regenerated by a model that has never seen the person. The output is a statistical guess consistent with pose, lighting and the prompt, which is why two tools produce two different bodies from one photo. Any “x-ray” claim is false.
Which model do undress apps use for inpainting?
Most do not disclose it. Observed behavior in our tests, such as 512 px native output and typical seam artifacts, matches Stable Diffusion 1.x or 2.x inpainting checkpoints fine-tuned on nude imagery, sometimes with LoRA adapters. A few services route requests to hosted model catalogs.
What do strength and guidance scale mean?
Strength sets how much of the masked area is regenerated: 1.0 replaces it fully, lower values keep part of the original. Guidance scale sets how strongly the prompt steers the result; the Diffusers default is 7.5. With the number of denoising steps, default 50, they control fidelity and speed.
Can inpainting be detected afterwards?
Sometimes. Seams, inconsistent noise patterns and lighting mismatches at the mask boundary are the usual tells, and forensic tools look for them. Provenance data such as C2PA Content Credentials record an edit when the software supports it, but a screenshot removes that record. No detector is conclusive.
Is inpainting legal to use on a photo of a person?
The technique is neutral; the subject and the result decide. Editing your own photo or a consenting adult’s is lawful. Producing an intimate image of an identifiable non-consenting person is a “digital forgery” under the TAKE IT DOWN Act and grounds for a civil claim under 15 U.S.C. §6851.