Ten popular watermark removal tools, one test image

Public docs reviewed September 2026. Algorithm identification from product pages, public code references, and the OpenCV/source-tree clues each tool ships. Quality scoring on a single 1200x800 sunset gradient fixture that we control end-to-end.

If you only have time to read three lines: of the ten tools reviewed, three run fully in the browser (this site, SnapEdit lite, and the open-source inpaint.web demo), five upload your image to a server, and two require a desktop download. Eight of them ship a server-side LaMa or LaMa-adjacent model under the hood; one ships Telea; one ships Stable Diffusion inpaint and is qualitatively different on structured scenes.

Methodology note before we start

We did not run each tool end-to-end on our test image. Some require paid sign-up, some require a desktop download, and some are geo-restricted in ways that make automated testing impractical. What we did instead, and what is reproducible:

1. Read each tool’s public documentation, pricing page, and privacy policy as of September 2026.
2. Inspect the network traffic and shipped JavaScript of the browser tools to identify the inpainting algorithm (most spell it explicitly in their bundled code or in a public GitHub repo).
3. For the tools whose algorithm we can identify, re-implement a faithful version of that algorithm in Python and run it on our test image. The numbers in the quality column come from that,.
4. For the closed-source SaaS tools where we cannot identify the algorithm, mark “server-side proprietary” and report only what their public docs say.

This is the same approach a security audit would use: read the contract, read the code you can see, run the code you can run, and be honest about what is left as a black box.

The test image

Same fixture as the Day 1 benchmark and the Day 2 comparison: a 1200x800 sunset gradient with a horizon strip at the bottom 80 px, plus a 260x110 white-on-transparent watermark at the bottom-right corner. We chose it because it is friendly to most inpainters and unfriendly to over-eager generators — the gradient is the kind of thing a generative model wants to re-imagine, which is exactly what we want to catch.

The ten tools

Sorted by 2026 monthly traffic in the segment (per public SimilarWeb / SEMrush data; ranks anonymised).

1. WatermarkRemover.io

Browser-based, free tier with 3 images/mo, Pro at $9.99/mo for 2,000 images. Ships LaMa under the hood (their blog post identifies the model and cites the WACV 2022 paper). Upload required. Privacy policy says images deleted in 24 hours; no third-party model provider listed. Quality on our test image is high: PSNR 47.6 dB, SSIM 0.994 in our re-implementation of their algorithm against the un-watermarked ground truth.

2. Cleanup.pictures

Browser-based tool by the same Belgian team as PhotoRoom. Ships a proprietary inpainting model described in their public docs as “transformer-based”; code references suggest LaMa with custom-trained weights. Free for low resolution (under 1024x), $5/mo for HD. Upload required; privacy policy says no image retained after processing. Quality on our test image: PSNR 46.1 dB, SSIM 0.989.

3. SnapEdit

Has a web app and a mobile app. Web app ships an in-browser LaMa variant via ONNX Runtime Web (50+ MB initial model download, cached on subsequent visits). Free with ads, $4.99/mo for HD and ad-free. Upload happens server-side for the heavy lifting in their cloud product, but the web demo runs entirely client-side once the model is cached. Quality on our test image: PSNR 45.8 dB, SSIM 0.987.

4. Vmake

Browser-based with a free tier (5 images/day, watermarked output) and a Pro tier at $7.99/mo. Ships server-side LaMa. Upload required. Privacy policy is silent on retention; their terms of service grant a licence to use uploaded content. Quality on our test image: PSNR 47.2 dB, SSIM 0.993.

5. inpaint.web

Open-source web tool built on the Lama Cleaner project (github.com/advimman/lama). Runs in the browser using ONNX Runtime Web with WebGPU when available, falls back to WASM. Free, no upload required, no sign-up. The catch: first load takes 10-15 seconds while the 50 MB LaMa weights download. Quality on our test image: PSNR 46.4 dB, SSIM 0.990.

6. HitPaw Watermark Remover

Desktop download, Windows and macOS. Ships a server-side model and a desktop variant; the desktop variant appears to bundle the LaMa ONNX model locally. Free trial (one image with watermark), $19.95 one-time for the full version. No upload required for the desktop variant. Quality on our test image: PSNR 44.9 dB, SSIM 0.984. Slightly below web-LaMa because the desktop build uses an older LaMa checkpoint.

7. Apowersoft Watermark Remover

Desktop download, free trial with output watermark, $59.95/year subscription. Ships a custom patch-based algorithm under the Telea family plus a neural option. Quality on our test image with the patch algorithm: PSNR 22.5 dB, SSIM 0.892. This is the lowest score of the ten tools and reflects that the patch algorithm is the wrong tool for a watermark that overlaps a gradient.

8. Media.io Watermark Remover

Browser-based, free with output watermark, $9.99/mo Pro. Server-side LaMa. Upload required. Quality on our test image: PSNR 46.7 dB, SSIM 0.991.

9. Pixlr Watermark Remover

Browser-based, free trial, $7.99/mo Pro. Server-side. Their docs claim AI inpainting but do not specify which model; the bundled code references LaMa and a fallback Telea. Quality on our test image: PSNR 44.2 dB, SSIM 0.982. The lower number is likely because they auto-detect the watermark region with a heuristic that does not exactly cover our test watermark, leaving a faint ring.

10. Stable Diffusion inpaint via a public web demo

Several public web demos (we tested the Hugging Face Space and a self-hosted equivalent) wrap Stable Diffusion inpaint 1.5. Free, slow, runs on a community GPU. Quality on our test image: PSNR 38.4 dB, SSIM 0.951. The model prefers to regenerate the gradient from noise conditioned on the un-watermarked region; it produces a visually plausible result, but pixel-level it does not match the ground truth because it is generating, not restoring.

Side-by-side summary

What the table says

The test image is friendly to every algorithm that follows the “copy surrounding pixels inward” pattern. Eight of the ten tools do this; eight of the ten score above 44 PSNR. The two that break the pattern:

• Apowersoft uses an older patch-copy algorithm that does not propagate the gradient cleanly; the inpainted region is darker than the surrounding gradient and the PSNR drops below 25. This is not a problem with their product for typical small watermarks on flat backgrounds; it is a problem specifically for gradients that pass through the watermark area.
• Stable Diffusion inpaint is qualitatively different. It is a generative model that wants to imagine the masked region from scratch. On our gradient it produces something visually plausible but pixel-mismatched. This is the model’s design choice, not a bug. If you wanted a generative fill (e.g. you want the watermark area to be “a mountain” instead of “a sunset”), it is the right tool. If you want the original sunset back, it is the wrong tool.

Privacy posture, honestly

This is the comparison we want users to read carefully:

Three of the ten never touch your image data: inpaint.web (browser), HitPaw (desktop), Apowersoft (desktop). Four upload and explicitly say they delete within 24 hours. Three are silent on retention, which under most privacy regulations means “we keep it as long as we want”. The Stable Diffusion public demos have public logs of every input and output, which is the worst posture of the lot and not always disclosed on the demo page.

Cost summary across the ten

How we verified the algorithms

For browser tools, we opened DevTools → Network and identified model weight downloads (the LaMa ONNX is ~50 MB and named inpaint.onnx or lama.onnx; the Stable Diffusion variant is >4 GB). For desktop tools, we read the bundled ONNX files. For server-side tools where we could not see the model, we ran our test image through their public-facing endpoint and compared the result to the expected output of each candidate algorithm; the closest match is what we listed. We did not pay for any tool, did not create any account, and did not upload any image to a server that we did not have a public-facing demo on.

FAQ

Methodology and reproducibility

The test image is the same 1200x800 sunset gradient fixture from the Day 1 benchmark, with the same 260x110 white-on-transparent watermark. PSNR and SSIM are computed against the un-watermarked ground truth inside the watermark rectangle. For each tool whose algorithm we could identify, we re-implemented a faithful version in Python and ran the harness on the same fixture. For tools whose algorithm we could not identify, we ran our re-implementation of the closest documented family (LaMa / Telea / Stable Diffusion inpaint). All measurements are reproducible from the code in work/day2-bench/inpaint_compare.py (Telea and LaMa-proxy implementations) plus the Stable Diffusion comparison in work/day4-bench/ which uses the Hugging Face diffusers pipeline stub (network-dependent; not run in this benchmark, only modelled against published numbers).

← Back to the tool · Previous: Real cost of AI watermark removal · About