How Undress Apps Work: Inpainting, Models, Limits
A plain description of what happens between an upload and an output, which components the reviewed tools appear to share, and why some photos fail.
An undress app is a hosted inpainting pipeline: it detects clothing regions in a photo, masks them and asks a diffusion model to synthesize replacement pixels. Across 15 catalog tools tested by September 3, 2026, the stages are the same; differences lie in models, input rules and queue design. Consent rules are in our Responsible AI policy.
What happens between upload and output?
Every tool we tested runs a variant of the same five-stage pipeline. None of the stages is unique to this category; each is a general computer-vision technique repurposed. Our inpainting entry covers the core method in more depth.
| Stage | What it does | Typical failure |
|---|---|---|
| Person and pose detection | Locates the body and estimates limb positions | Group photos, unusual angles, partial bodies |
| Clothing segmentation | Produces a pixel mask of garment regions | Patterned fabric, loose layers, hands over clothing |
| Diffusion inpainting | Regenerates masked pixels conditioned on the unmasked image and a text or pose prompt | Edge seams, mismatched skin tone, anatomical errors |
| Upscaling and blending | Raises resolution and smooths seams | Plastic texture, halo around the mask |
| Post-filters | Watermarks, safety classifiers, age checks | Missing on cheaper tools |
The provider decides how much compute to spend at stage 3. That decision, more than any “proprietary model” claim, explains most differences in output and price.
Which models power these tools?
Public evidence points to fine-tuned latent diffusion models, the family described in the 2022 Rombach et al. paper that underlies Stable Diffusion. Providers add three things: a segmentation network to draw the mask, a small fine-tune or adapter trained on the target domain, and a conditioning network for pose so the output follows the original body position. Wavespeed exposes this structure directly, selling a “clothes-changer” model at $0.12 per run alongside other models from $0.07 and video at $0.10 per second through an API. Consumer tools hide the same components behind credits: Ainudez charges 15 credits per image and 50 to 300 for short video, Deep Undress 10 stars per image and 25 per video. Our deepfake entry explains how this differs from face-swap methods.
Our LDMs achieve new state of the art scores for image inpainting and class-conditional image synthesis and highly competitive performance on various tasks … while significantly reducing computational requirements compared to pixel-based DMs.
— Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, accessed Sep 3, 2026
Why do results fail on some photos?
Failures cluster around the mask and the source. Deep Undress produced edge artifacts on low-light sources in our August 14, 2026 test; Undresswith AI needs sources of at least 1280 pixels, and smaller inputs lose detail after upscaling; UndressHer’s daily watermarked token yields softer output than paid tokens, which suggests a cheaper sampling setting. Patterned or layered clothing confuses segmentation, and any photo with more than one person raises error rates. The quality sub-score in How we rate is 25 percent of the total precisely because marketing samples never show these cases; our Editorial Test Notes record them.
| Tool | Input rule or limit observed | Output notes | Test date |
|---|---|---|---|
| Ainudez | 15 credits per image; video 480p 5 s to 720p 7 s | Quality 9.5, highest in the catalog | Aug 27, 2026 |
| Pornworks | Registered tier queues without a cap; guests 1 watermarked image per day | Quality 9.4; video and face swap only on the Ultimate plan | Aug 12, 2026 |
| Deep Undress | 10 stars per image, 25 per video | Speed 9.7; edge artifacts on low-light sources | Aug 14, 2026 |
| Undresswith AI | Sources of 1280 px or larger; HD unlock $0.99 | Detail loss below the threshold | Aug 28, 2026 |
| Wavespeed | $0.12 per run; API access | Outputs expire within 7 days | Aug 16, 2026 |
| UndressHer | 1 token per image; daily token watermarked | Softer output on the daily token | Aug 17, 2026 |
What do queues and credits reveal about the backend?
Speed is 15 percent of our score because it exposes infrastructure choices. A slow queue for a no-cost tier, as at Pornworks, means shared GPUs with paid jobs prioritized; the stalls we logged are the visible side of that. Per-run pricing with credits that do not expire, as at Wavespeed, indicates metered cloud compute passed through to the customer. Video costing 20 times an image at Ainudez reflects frame-by-frame inpainting plus temporal smoothing. The clothes-remover category page compares speeds side by side; the pattern is that price per image and wait time move in opposite directions.
Where are the limits set by law and policy?
The technology has no built-in notion of consent, so limits arrive from outside. Under the EU AI Act, Article 50 requires that AI-generated or manipulated images be marked as such in a machine-readable way from August 2, 2026; providers serving EU users must label outputs. In the United States, the TAKE IT DOWN Act, enforced by the FTC since May 19, 2026, obliges platforms to remove non-consensual intimate images within 48 hours, and Minnesota bars access to the tools outright from August 1, 2026. The reviewed tools respond with age gates, consent clauses and geo-blocks; our undress-AI term entry tracks how these constraints are changing the category, and the state laws entry lists jurisdictions.
Sources and verification
Primary sources checked for this article. Service facts come from the linked editorial reviews and their dated test notes.
- Rombach et al. — High-Resolution Image Synthesis with Latent Diffusion Models (arXiv, 2022) — accessed Sep 3, 2026
- Regulation (EU) 2024/1689 (AI Act), Article 50 transparency obligations — accessed Sep 3, 2026
- FTC — Complying With the Take It Down Act — accessed Sep 3, 2026
- Minnesota House — New law bans access to nudification technology — accessed Sep 3, 2026
Consent and reporting
Related reading
Inpainting explained
How diffusion models fill masked regions and where the method breaks.
Read the entry →Undress AI: term entry
Definition, history and current legal status of the category.
Read the entry →Clothes-remover tools compared
Speed, quality and price side by side from dated tests.
See the comparison →Deepfake explained
How face-swap and inpainting methods differ and why the law treats them alike.
Read the entry →Market overview: State of Undress Apps – Annual Editorial Report · Undress Apps Directory: 15 Reviewed Tools, Filters, Scores.
Frequently asked questions
Do undress apps use a special AI model?
No. Public evidence and the tools’ own API documentation point to fine-tuned latent diffusion models, the same family behind Stable Diffusion, combined with a segmentation network and a pose-conditioning network. Providers differentiate on training data, compute budget and interface, not on a fundamentally different method.
Why does the same photo give different results in different apps?
The segmentation mask, the fine-tune, the sampling steps and the upscaler all differ. A tool spending more compute per image, or requiring larger inputs as Undresswith AI does, produces fewer seams. Cheaper tiers often use fewer steps, which is why UndressHer’s daily token output is softer than its paid output.
Why do some tools need photos of 1280 pixels or more?
Inpainting works in a compressed latent space and the upscaler cannot invent detail that the source lacked. Undresswith AI sets a 1280-pixel minimum for that reason; below it, edges blur after upscaling. Larger inputs also give the segmentation network more pixels to separate fabric from skin, reducing mask errors.
How long does an image take to generate?
From a few seconds to several minutes, depending on queue priority. Deep Undress scored 9.7 for speed in our test; Pornworks’ registered no-cost tier stalled in a shared queue while paid jobs ran first. Video takes far longer: Ainudez prices a 5-second 480p clip at 50 credits and a 7-second 720p clip at 300.
Can the output be detected as AI-generated?
Often, but not reliably by eye. Seams, texture and lighting mismatches are common tells, and EU AI Act Article 50 requires machine-readable marking of synthetic images from August 2, 2026. Several tools add watermarks on no-cost tiers only. Detection tools exist but produce false results, so provenance metadata is the more dependable signal.
Does understanding the technology make using it legal?
No. Legality depends on consent and jurisdiction, not on method. Processing a photo of anyone other than yourself or a consenting adult can violate state law, and publishing the result is a federal offense under the TAKE IT DOWN Act. Minnesota bars access to such tools entirely from August 1, 2026.