Is this AI image upscaler really free and private?

Yes. There is no account, watermark or limit on the number of images. The upscaling runs on your own computer or phone, and your image is never uploaded to a server — only the AI model is downloaded, once.

How long does upscaling take?

With WebGPU, the Fast model upscales a typical 1000-pixel image in a few seconds. On the CPU it can take a minute or more, and the Quality model is about ten times slower again. The progress bar shows the elapsed time and an estimate of the time left after the first tile.

Why does the first upscale take longer?

The first run downloads the AI model — 4.6 MB for Fast or 64 MB for Quality — and prepares it for your graphics card or CPU. The model is then cached in your browser, so later runs skip the download.

Can AI upscaling restore a blurry photo?

It can make a low-resolution or mildly blurred photo look much sharper, and it cleans up JPEG artifacts and noise. It cannot recover detail that was never captured: tiny faces, unreadable text and strong motion blur stay ambiguous, and the network may invent texture there.

What is the largest image I can upscale?

The result is limited to about 36 megapixels on a computer — for example 6000 × 6000 pixels — and about 16 megapixels on phones and tablets. If you ask for more, the input is reduced first so the result fits, and the tool tells you.

Does it keep transparent backgrounds?

Yes. Save as PNG or WebP and the transparency is preserved: the alpha channel is enlarged with bicubic resampling alongside the AI-upscaled colours. JPEG has no transparency, so those areas become white.

My browser says WebGPU is not available. Will it still work?

Yes, it falls back to WebAssembly on the CPU, which works in every modern browser but is slower. For the best speed use a recent version of Chrome, Edge or Safari on a computer with a graphics card.

Which model is used, and can I use the results commercially?

The tool uses the official Real-ESRGAN weights (general v3 and x4plus), released under the permissive BSD-3-Clause license. The images you create are yours to use; the license covers the model, not your pictures.

AI Image Upscaler: Enlarge Photos 2× or 4×

Upscale images 2× or 4× with AI — on your own device

This AI image upscaler enlarges small, soft or low-resolution pictures with <strong>Real-ESRGAN</strong>, a neural network trained to rebuild the edges, textures and fine detail that ordinary resizing smears away. It runs entirely in your browser — on your graphics card through WebGPU when available, otherwise on the CPU through WebAssembly. Your image is never uploaded: the only thing downloaded is the AI model itself, once, after which it is cached on your device.

How to upscale an image

1Drop an image on the page, choose one, or paste it with Ctrl/⌘ + V. PNG, JPEG, WebP and AVIF all work.
2Pick a model: Fast (a 4.6 MB download, good for most pictures and screenshots) or Quality (64 MB, the best detail on photos).
3Choose 2× or 4× and press Upscale. The first run downloads the model; progress is shown tile by tile with the time remaining.
4Drag the slider across the image to compare before and after, and zoom in to inspect the detail.
5Choose PNG, JPEG or WebP and press Download — or send the result straight to another tool, such as the image resizer.

Why AI upscaling looks sharper than resizing

A normal resize — bilinear, bicubic or Lanczos — can only spread the pixels you already have over a larger grid. It invents no detail, so a 4× enlargement looks soft and blocky: edges blur, text turns fuzzy and JPEG artifacts grow with the image.

A super-resolution network works differently. Real-ESRGAN was trained on millions of pairs of sharp images and deliberately degraded copies — blurred, downscaled, noisy and JPEG-compressed — and learned to map the degraded version back to the sharp one. Given your picture, it predicts plausible high-frequency detail: crisp edges, hair, foliage, fabric and lettering. That is also why it removes compression blocks and mild noise as a side effect.

The detail it adds is a prediction, not a recovery. It cannot reveal a license plate that was never legible or a face that was a few pixels wide, and on very small or heavily compressed images it can look painterly. Compare with the slider and zoom to 100% before you rely on the result.

Fast or Quality: which model to choose

Fast — Real-ESRGAN general v3: <code>realesr-general-x4v3</code>, a compact SRVGGNet with about 1.2 million parameters and a 4.6 MB download. It is roughly ten times quicker than the Quality model and handles illustrations, screenshots, game art, product shots and most photos well. Start here.
Quality — Real-ESRGAN x4plus: <code>RealESRGAN_x4plus</code>, the original RRDBNet with 16.7 million parameters and a 64 MB download. It restores more texture in real photographs — skin, foliage, architecture — at the cost of time: on a CPU, a large photo can take several minutes. A graphics card with WebGPU makes it practical.
2× versus 4×: Both models are 4× networks. For 2× the tool runs the 4× model and then averages every 2 × 2 block of the result, which is noticeably sharper than enlarging 2× directly and keeps memory lower, because the full 4× image never exists.

How it runs in your browser

The models run on ONNX Runtime Web in a background worker, so the page stays responsive. When your browser supports WebGPU — recent Chrome, Edge and Safari on most computers — the network runs on your graphics card. Otherwise it falls back to WebAssembly on the CPU, which is slower but works everywhere. The panel shows which one is in use.

Large images are split into overlapping tiles of 128 to 256 pixels. Each tile is upscaled with 16 pixels of surrounding context and only its centre is kept, so there are no visible seams while the memory needed stays small. That is also what makes the progress bar, the time estimate and the Cancel button possible.

The model file is downloaded from Hugging Face the first time you use it and kept in your browser's cache storage, so later visits start almost instantly and work offline. Your images themselves are only ever read and processed on your device.

Transparency, output formats and size limits

Transparent PNGs keep their transparency: the network enlarges the colours, and the alpha channel is enlarged separately with bicubic resampling and recombined.
PNG is lossless and keeps transparency — the best choice for further editing. JPEG and WebP are much smaller; JPEG fills transparent areas with white.
To protect your browser from running out of memory, results are capped at about 36 megapixels on computers and 16 megapixels on phones and tablets. A larger request reduces the input first and tells you so.
Need an exact size afterwards? Send the result to the image resizer, convert it with the image converter, or cut out the subject with the background remover.

Credits and license

Upscaling uses Real-ESRGAN by Xintao Wang, Liangbin Xie, Chao Dong and Ying Shan (“Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data”, ICCV Workshops 2021), released under the BSD-3-Clause license. The ONNX conversions are the BSD-3-Clause jonathanst29/tinier-upscale-models exports on Hugging Face, pinned to a fixed revision. Inference is powered by Microsoft's MIT-licensed ONNX Runtime Web.

Frequently asked questions

Is this AI image upscaler really free and private?: Yes. There is no account, watermark or limit on the number of images. The upscaling runs on your own computer or phone, and your image is never uploaded to a server — only the AI model is downloaded, once.
How long does upscaling take?: With WebGPU, the Fast model upscales a typical 1000-pixel image in a few seconds. On the CPU it can take a minute or more, and the Quality model is about ten times slower again. The progress bar shows the elapsed time and an estimate of the time left after the first tile.
Why does the first upscale take longer?: The first run downloads the AI model — 4.6 MB for Fast or 64 MB for Quality — and prepares it for your graphics card or CPU. The model is then cached in your browser, so later runs skip the download.
Can AI upscaling restore a blurry photo?: It can make a low-resolution or mildly blurred photo look much sharper, and it cleans up JPEG artifacts and noise. It cannot recover detail that was never captured: tiny faces, unreadable text and strong motion blur stay ambiguous, and the network may invent texture there.
What is the largest image I can upscale?: The result is limited to about 36 megapixels on a computer — for example 6000 × 6000 pixels — and about 16 megapixels on phones and tablets. If you ask for more, the input is reduced first so the result fits, and the tool tells you.
Does it keep transparent backgrounds?: Yes. Save as PNG or WebP and the transparency is preserved: the alpha channel is enlarged with bicubic resampling alongside the AI-upscaled colours. JPEG has no transparency, so those areas become white.
My browser says WebGPU is not available. Will it still work?: Yes, it falls back to WebAssembly on the CPU, which works in every modern browser but is slower. For the best speed use a recent version of Chrome, Edge or Safari on a computer with a graphics card.
Which model is used, and can I use the results commercially?: The tool uses the official Real-ESRGAN weights (general v3 and x4plus), released under the permissive BSD-3-Clause license. The images you create are yours to use; the license covers the model, not your pictures.