iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
ClearPix is designed to remove watermarks in the browser rather than send the image or video to a processing server. In a 2026 engineering account, its developer describes a static-site architecture that downloads AI models to the browser, then decodes, edits, and re-encodes media in the user’s tab. That can keep media off a processing backend, but it does not mean the app makes no network requests: model files, runtime assets, and aggregate analytics still travel over the network. The privacy and performance details below are the developer’s account, not an independent audit or device test. Source: Umasou, DEV Community, October 1, 2026.
What “without uploading” means in this editor
In the architecture described by ClearPix’s developer, the browser downloads the editor’s runtime and model weights, then performs the media work locally. The user’s image or video is decoded, processed, and re-encoded in the open tab; the system is described as having no processing backend to which it sends that media.
That is a narrower and more useful privacy claim than “no network traffic.” The application still needs to retrieve software and model data, and the author says it collects aggregate analytics. Umasou describes those events as contractually barred from containing images, filenames, or file contents. The article also says media-bearing canvas, image, and video elements are masked in Microsoft Clarity session replays. These are the author’s descriptions of controls and data handling, not findings from an independent inspection of network requests or a privacy audit. Umasou’s technical account.
How the browser-based editing pipeline works
1. The browser obtains the runtime and model
Rather than sending a file to a server for inference, the page downloads ONNX Runtime Web and the model weights needed by the editor. The author says the weights may come from storage or CDN sources and are cached in IndexedDB so they do not have to be fetched from scratch on every use.
#1 Best Overall
- Exceptional Performance and Productivity: Experience smooth and responsive performance powered by an AMD Ryzen 7 7730U processor and 16GB memory and 512GB SSD. Enjoy extended productivity thanks to exceptional battery life and the support of Copilot, your everyday AI companion.
- Copilot in Windows - your AI Assistant: Do more, quicker than ever across multiple applications with the centralized generative AI assistance of Copilot in Windows Accessible with a single touch of the Copilot Key
- Immersive Visuals: With its narrow bezel design the 15.6" 1080p Full HD IPS display is perfect for casual web browsing and watching movies or streaming, allowing for a sharp, detailed view of what's in front of you. And with Acer BluelightShield, lower the levels of blue light to lessen the negative effects of blue light exposure.
- User-Friendly by Design: Seamlessly connect or charge your devices through a full-function USB Type-C port, while Wi-Fi 6 and HDMI 2.1 connectivity enhance your digital experiences to be faster, smoother, and more enjoyable.
- Unlock More with AcerSense: Intuitive device control is available at the touch of a button with AcerSense, which manages battery life, storage, and apps for optimal performance. Acer TNR solution and Acer PurifiedVoice enhance your video calling experience to a new level of clarity and quality.
2. The runtime selects an inference path
The described implementation prefers WebGPU for model inference and falls back to WebAssembly (WASM) when WebGPU is unavailable or unreliable. The fallback is important because browser and device support is uneven: local inference depends on the user’s browser, available memory, GPU behavior, and runtime environment.
3. Media is processed in the tab
The editor decodes the selected media, applies the model to the marked region, and re-encodes the result in the browser. In this design, the media bytes are not supposed to go to a processing server. That distinction does not by itself prove that a live deployment or every third-party script behaves as described; the cited article reports the developer’s design and stated controls.
4. Model data is checked and cached
Model files are large, so the account describes pinned SHA-256 hashes to verify downloads and cached data. The developer also describes chunked IndexedDB writes, retries against mirrors when downloads stall, and fallback model specifications. These measures aim to handle interrupted delivery and avoid preserving corrupted model data, but do not remove the initial download or local storage requirements. Implementation details from Umasou.
Free tools Windows power users keep installed
One-click scans. No signup required.
What the reported performance numbers show
The figures below are ClearPix team measurements and engineering observations reported by Umasou in 2026. They describe this implementation under the team’s conditions, not universal browser benchmarks or guaranteed results on a reader’s device.
Rank #2
- [Powerful Performance] 15th Gen Ultra 7 265U 2.10GHz Processor (upto 5.3 GHz, 12MB Cache, 12-Cores, 14-Threads, 2 Performance-cores); Intel Graphics Integrated Graphics
- [High Speed and Multitasking] 32GB DDR5 SODIMM; Bluetooth 5.4, RJ-45, Webcam, Fingerprint, Backlit KB Standard Keyboard
- [Superior Display] 14.0" IPS WUXGA (1920x1200) 60Hz Display; 65W PSU, Type-C Power-In, 3-Cell 62 WHr Battery; Meteor Silver Color
- [Enormous Storage] 1TB PCIe NVMe SSD; 1 x HDMI 2.1, 2 Thunderbolt 4, USB 3.2 Type-C, Headphone/Microphone Combo Jack
- Windows 11 Pro-64, 1 Year Manufacturer warranty from MichaelElectronics2 (Professionally upgraded by MichaelElectronics2)
| Reported item | ClearPix figure | What it describes |
|---|---|---|
| Inpainting time per region | 2,100–3,500 ms, falling to roughly 400 ms | The team’s reported comparison after enabling the WebGPU execution provider. |
| Model delivery rate | 75–110 KB/s for direct r2.dev URLs; 2.3 MB/s for a same-zone custom domain | The team’s measured delivery comparison for its model distribution setup. |
| WASM dependency size | About 28 MB | The article’s stated size for the jsep.wasm dependency. |
| Large model and storage writes | 198 MB model; writes above 64 MB chunked into 32 MB pieces | The engineering account says a single structured clone of the model caused renderer out-of-memory trouble. |
| Cached-model hashing | About 100 ms for a 67 MB model | The team’s reported hash-check measurement. |
| Release archive size | 27.4 MB reduced to 10 MB | The reported archive change after excluding model and benchmark assets. |
The numbers illustrate the central engineering trade-off: keeping inference in the browser avoids routing media through a processing service, but requires careful handling of model delivery, browser memory, storage, and compute. They should not be used to predict how quickly another browser-based editor will run. The figures are attributed to the ClearPix team in Umasou’s article.
Why local processing is slower or less predictable at first
First use has extra work
The developer says a first visit takes longer because model files must download and WebGPU shaders must compile. Caching may reduce repeated model downloads, but it cannot make the initial transfer disappear.
Devices do not behave like a data-center GPU
Server hardware is controlled; a browser editor runs on a wide range of consumer hardware. The article notes that WebGPU can fail or stall and mobile devices have constrained memory. A model that works well on one machine may run more slowly or hit resource limits on another.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Fallbacks mitigate some failures, not all
The described implementation uses a watchdog and WASM fallback for some WebGPU problems. WASM performance can itself be constrained: the author says missing cross-origin isolation can limit threading. A fallback path improves resilience, but it is not evidence that every browser and device will work reliably.
Rank #3
- 【Snapdragon X X1-26-100 and 45 TOPS AI for Professionals】 Experience the peak of 8-Core performance up to 3.8 GHz. This HP 16" Business Laptop uses massive AI power to automate complex workflows, making it a reliable tool for college students and business users managing specialized professional software.
- 【Stunning FHD+ IPS Panel for Visual Clarity】 The 16:10 aspect ratio on this FHD Laptop provides an immersive viewing experience. It is perfect for business users analyzing large spreadsheets and students who need a clear screen for their daily college lectures and research.
- 【AMD Radeon 740M Graphics for Creative Projects】 Enhance your visuals with integrated Radeon 740M graphics. This HP Laptop provides the necessary power for traditional video editing and high-definition media, perfect for creators who require smooth performance for their digital standard of work.
- 【16GB DDR5 RAM and 512GB SSD for High-Speed Multitasking】 Featuring 16GB LPDDR5x memory at 8448 megahertz, this Computer Laptop handles heavy apps with ease. The 512GB SSD provides ample storage for large data files, making it a fast and reliable choice for professional and educational use.
- 【Copilot+ PC Experiences for Modern Efficiency】 Unlock a smarter way to work with advanced AI assistance for daily tasks. This computer notebook netbook simplifies your digital life, helping business users and students organize schedules and notes with intelligent, high-speed features.
Model delivery and cache integrity matter
Large weights increase first-run download time and storage pressure. ClearPix’s account describes chunking large writes, retrying mirrors after stalls, and checking cached bytes against pinned hashes. Those are safeguards for this particular implementation, not a general guarantee that browser storage is unlimited or that downloads cannot fail. Umasou discusses these constraints and mitigations.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to assess a watermark remover’s privacy claim
“Runs in your browser” is a claim about where processing is intended to happen; the meaningful question is what data actually leaves the device. For any editor, look for clear answers to these points:
- Media destination: Does the service say explicitly whether the image or video is sent to a processing backend?
- Other network traffic: Does it distinguish the media from model downloads, runtime files, analytics, and third-party scripts?
- Media-bearing page content: Are canvases, previews, and video elements protected from session replay or similar telemetry?
- First-use cost: How large are the models, and what must download before the editor works?
- Local reliability: What happens when WebGPU is missing, fails, or stalls, and does the fallback suit the device?
- Inspectable evidence: Can the architecture and privacy statements be checked against observable network behavior and the service’s policies?
Umasou’s article offers an implementation account for ClearPix, including stated analytics limits and replay masking, but it does not independently verify the deployed application’s traffic. It also does not provide a comparative survey establishing how other watermark removers handle uploads, so it cannot support a broad claim about the whole market.
The trade-off: less server exposure, more work on your device
Client-side processing can reduce the need to send media to a server, which may matter for private or sensitive images. In exchange, the user’s browser and hardware take on the compute and reliability burden. The editor must deliver sizable model assets, fit them into browser storage and memory, and cope with variable GPU and WASM support. Server processing moves more of that compute burden to the service, but requires the user to evaluate where uploaded media is handled and retained. The ClearPix account documents one attempt to make that trade-off work; its reported measurements and privacy assurances should be read as claims about that implementation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
