ByteScope

AI Image Upscaler

Make a small picture bigger with an AI model that runs in this tab — drop it in, pick 2x or 4x, drag to compare against plain stretching, and download. Nothing is uploaded.

Your files never leave your browser — all processing is local.

The picture

Drop an image here

or click to browse — you can paste one too. JPEG, PNG or WebP.

Upscale

Scale

Save as

PNG is usually much larger than JPEG, so the size above can be an underestimate.

Uses Real-ESRGAN compact 4×, 4.9 MB, BSD-3-Clause. It is downloaded once on the first run. 2× is that 4× result scaled down.

Acceleration: Checking what this browser can do…

What to expect

  • AI upscaling guesses the detail, it does not recover it. Text, number plates and faces can be invented, so do not use the result as evidence or to identify anyone.
  • The model is trained on ordinary photos. Illustrations, anime and pixel art can come out with softer lines or a painted look.
  • A heavily compressed JPEG has its compression noise enlarged along with the picture.
  • A transparent background becomes white: the image is placed on white before upscaling, so the result has no transparency.
  • Input is limited to about 0.84 megapixels because the model always works at 4×. A bigger image is refused before it starts.
  • Without WebGPU this runs on the CPU and is much slower. The page shows which path it is using.
  • The picture is never uploaded. Only the model file is fetched, once.

About this tool

Stretching a small picture in an editor smears every pixel into its neighbours, which is why the result looks soft. This page is an AI image upscaler that runs Real-ESRGAN compact, a 4.6 MB super-resolution model released under the BSD-3-Clause licence, on your own machine: it looks at small tiles of the picture and draws plausible edges and texture at four times the size. Choose 2x or 4x, and the page tells you the output size and a rough file size before you start, so you never wait for something that will not fit. It is for anyone who needs to enhance the resolution of a logo, a screenshot or a product shot without handing the file to someone else's server.

What the model really does

Upscaling is guessing, not recovering. The model has never seen your original scene; it fills in detail that looks right for the kind of picture it was trained on. On ordinary photos and screenshots that is a clear improvement over stretching. On text, number plates and faces the invented detail can be wrong in ways that look convincing, so do not use an enlarged picture as evidence or to identify anyone. Illustrations, anime and pixel art can come out with softer lines or a painted look, and a heavily compressed JPEG has its blocky noise enlarged along with everything else. The compare slider has a plain-stretch mode and a raw-pixel mode so you can see exactly what the model added.

Size limits, speed and privacy

The model always works at four times the input, so the limit is on the input size: about 0.84 megapixels in, roughly a 915 by 915 picture. A bigger image is refused before anything runs, and the page offers to shrink it to the largest size that fits, then upscale that. With WebGPU the work runs on the graphics card; without it the page falls back to the CPU through WebAssembly, which works but is much slower, and it links to the computer specs page to show why. The picture is read and processed in this tab and never uploaded; only the model file is fetched, once, and it is shared with the photo restoration page. Output is a fresh PNG, JPEG or WebP at quality 0.92, with EXIF orientation applied and other metadata left out. For a faded old print that also needs repair or colour, use the photo restoration tool instead.

Frequently asked questions

Does it upscale an image without losing quality?

Compared with a plain resize, yes for most photos and screenshots: a normal resize only interpolates between existing pixels, so edges blur, while the model redraws edges and texture and the same picture looks sharper at the larger size. But sharper is not the same as lossless — the extra detail is the model's guess, not pixels recovered from the original. The compare slider has a plain-stretch mode so you can judge it on your own image. For flat artwork or pixel art the gain is smaller and sometimes negative.

Why is there no 8x option?

Running the model twice enlarges its own invented detail and noise, so the second pass looks less honest, not better. The page stops at 4x. If you need a bigger file, upscale at 4x and check the result before you commit to going further.

Why was my image refused?

The model always runs at four times the input, and the browser has to hold that whole result in memory and draw it on one canvas. The page therefore limits the input to about 0.84 megapixels (roughly 915 by 915), and it tells you the largest longest-side that fits for your picture's shape. A button shrinks the picture to that size first, then upscales it, so a large phone photo can still be used, at the cost of starting from fewer pixels.

Can I trust the text and faces in the result?

No. The model guesses the detail, and a guess at a face, a number plate or small print can look sharp and still be wrong. Use the output for presentations, prints and social posts, not as evidence and not to identify a person or read a value you could not read in the original.

Why is it slow on my computer?

Without WebGPU the model runs on the CPU through WebAssembly, which can take many times longer than on a graphics card. The page shows which path it is using and links to the computer specs page, which checks what your browser exposes. Closing other heavy tabs helps, and a smaller input always runs faster.

Are my pictures uploaded?

No. The picture is decoded, upscaled and encoded inside this tab; there is no upload endpoint. The only download is the 4.6 MB model file, fetched once and kept in the browser cache, shared with the photo restoration page, so it is not fetched again there.

Related tools