AI Image Upscaler
Upscale images 2x or 4x without sending them to a server. Two Swin2SR models load directly in your browser — pick classical 2x for clean line art and screenshots, or real-world 4x for noisy photos.
Why an AI Tool That Runs In Your Browser
Server-based upscalers see every image you send. For an unreleased product render, a screenshot of a private dashboard, a watermarked client mockup, or any visual you cannot upload to a public service, that is a real problem. This tool loads Swin2SR — a small super-resolution model from CAIDAS at the University of Wuerzburg — and runs every upscale directly in your browser. The image you pick never travels anywhere. No request log, no retention policy, no upscaled copies stored on a server you cannot audit. Two model variants are available in one dropdown so you can match the input. Classical 2x is the cleanest pick for line art, vector exports rasterized at low resolution, UI screenshots, logos, and any input that started life as a sharp source. Real-world 4x is trained on compressed JPEGs and noisy photographs, so it handles the kinds of images you actually find in the wild — phone photos, downsampled stock, old web images, social-network compressed art. Both checkpoints are released by CAIDAS under the Apache 2.0 license, which permits commercial use without strings.
How AI Image Upscaler Works
Pick the 2x classical model or the 4x real-world model, then click Load model. The browser downloads the ONNX weights — about 22 MB for either variant — and caches them in IndexedDB so future visits skip the download. Drop or pick an image from your device. The tool resizes anything larger than 1024 pixels on the longest side before inference to keep memory bounded; if you need higher input resolutions, downscale in another tool first. Click Upscale. On a modern laptop with WebGPU the inference runs in two to ten seconds depending on input size; the WebAssembly fallback adds roughly a 3x multiplier. The output appears below the input with a download button — saved as PNG, lossless. Swin2SR works best on images with clear structure and recoverable detail. It is not magic: a 64-pixel thumbnail will not become a poster, and heavily compressed inputs will retain some of their compression artefacts even after upscaling. For comparison, line art and screenshots usually look noticeably sharper at 2x; photographs gain more from the real-world 4x model since it was trained specifically on degraded photographic inputs.
Frequently Asked Questions
What is the file size for each Swin2SR variant?+
Each Swin2SR variant is about 22 MB quantized, downloaded once from the Hugging Face CDN. The browser caches the weights in IndexedDB so later visits load in a few seconds. If you switch between the 2x and 4x variants the second model also downloads once and is then cached separately.
Does my image leave my device during upscaling?+
No. After the model finishes downloading on first use, every upscale runs entirely in your browser. The image you pick stays on your device and is never sent to our servers or to the Hugging Face CDN.
When should I use the 2x classical model vs the 4x real-world model?+
The 2x option uses Swin2SR classical, trained on bicubic-downsampled inputs with no extra noise. It is the cleanest pick for line art, UI screenshots, logos, and other sharp-source images. The 4x option uses Swin2SR real-world, trained with the BSRGAN degradation pipeline to handle JPEG compression and noise. It is the better pick for phone photos, social-media images, and old web graphics.
What licenses do the upscaler models use?+
Both Swin2SR variants are released by CAIDAS (the Computer Vision Lab at the University of Wuerzburg) under the Apache 2.0 license. Apache 2.0 permits commercial use, modification, and redistribution. The ONNX weights are mirrored by the Xenova organization on Hugging Face under the same license.
What input image sizes does the upscaler accept?+
Any common format (PNG, JPEG, WebP, GIF still frames). To keep browser memory bounded, inputs larger than 1024 pixels on the longest side are resized before inference. If you need to upscale a larger source, downscale to roughly 1024 pixels first, run the upscale, then run it a second time on the result if you need more resolution.
Will upscaling fix a blurry or compressed photo?+
Partially. The real-world 4x model recovers some detail from compressed photos but cannot invent information that the source genuinely lost. Heavy JPEG artefacts, severe motion blur, and very small thumbnails are out of reach for any upscaler, including this one. The result will be cleaner and sharper but not equivalent to a higher-resolution original.
Does the upscaler work on phones?+
Yes on modern iPhones (iOS 17 or later) and recent Android phones, though performance is slower than on a laptop. Memory pressure is the practical limit on phones — inputs above about 768 pixels on the longest side may fail on older devices. WebAssembly fallback works everywhere; WebGPU acceleration kicks in automatically on supported browsers.
Why does the upscaled image look slightly different in color?+
Swin2SR runs on normalized RGB tensors and writes the result back into a fresh canvas. Tiny color shifts of a few percent are normal and come from the float-to-byte rounding at the output stage. If exact color fidelity is critical for your workflow, run a color comparison after upscaling and apply a correction layer in your editor of choice.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service