You have a small image, say a profile photo, an old scan or a logo someone sent at 300 pixels wide, and you need it bigger. Stretching it in a document makes it blurry. The Image Upscaler offers two ways to enlarge it, and picking the right one saves a lot of waiting and disappointment. This guide explains both, the settings, the size limits and what you should expect.
The two methods
The Method control has two options:
- Sharp resize is a fast, non-AI method: Lanczos enlarging followed by gentle edge-aware sharpening. It takes about a second and needs no download. It is the default on browsers without WebGPU.
- AI uses a neural network that adds believable detail. The model is downloaded once (about 5 MB) and then works offline. With the AI method you also pick an Image type: Photo for photos and screenshots, or Anime / art for drawings.
Everything runs on your own device and nothing is uploaded.
Which one should you choose?
| Situation | Pick |
|---|---|
| You need a quick, clean enlargement and the source is already decent | Sharp resize |
| Soft, noisy or heavily compressed photo and you can wait | AI, Photo |
| Cartoon, illustration or anime artwork | AI, Anime / art |
| Phone or a browser without WebGPU, large image | Sharp resize, or a smaller image for AI |
| Many images in a batch | Sharp resize first; AI only on the ones that matter |
Speed: why the browser matters
The AI uses WebGPU when your browser has it, which today means Chrome and Edge on a computer. That path is roughly ten times faster than the fallback. Without WebGPU, the AI runs on the CPU, one thread at a time, and a large image can take minutes. The tool shows a warning with an estimate when the job will take long, and recommends a smaller image or a different browser. Safari, Firefox and many phones usually fall into the slow case, so Sharp resize is the sensible choice there.
Step by step
- Drop one image or several. JPG, PNG and WebP are accepted, up to 50 MB each.
- Choose the Method, and for AI the Image type.
- Choose Enlarge: 2x or 4x.
- Choose Save as: PNG, JPG or WebP.
- Press the upscale button. The first AI run downloads the model; later runs start straight away.
- Drag the comparison slider between Before and After and look at edges, text and faces.
- Download the result, or all results as one ZIP when you did a batch.
Size limits
The limit is on the size of the result, not just the input. According to the tool's FAQ, the result can be up to 64 megapixels on a computer, 32 on a computer with little memory, and 16 on a phone (8 on an older one). A 2000 x 1500 photo at 4x makes 48 megapixels, which fits a normal computer but not a phone. If an image is too big for 4x, the tool asks you to use 2x. Plan ahead by choosing the smallest enlargement that fits your needs.
What upscaling cannot do
- It cannot recover what was never captured. AI adds believable detail and cleans soft edges and JPEG blocks, but it does not reveal the actual small text or the real face of a blurry crop. Faces and fine text can look slightly different from the original, so check the result before using it for anything important such as ID or documents.
- It does not fix motion blur or heavy out-of-focus areas. The result will look cleaner but not truly sharp.
- Very smooth gradients may gain faint texture with AI, because it is built to look sharp rather than to match pixel for pixel.
Transparency and formats
PNG and WebP keep transparency; the alpha channel is enlarged smoothly. JPG cannot hold transparency, so transparent areas turn white. Save as PNG when you plan to edit further, WebP for a smaller file on the web, and JPG for maximum compatibility with photos.
After upscaling
Enlarged images are heavy. Reduce the file with the Image Compressor before you share it, or set exact dimensions with the Image Resizer if you overshot. If the subject needs a clean cut-out, use the Background Remover after upscaling, because the cut-out works on the larger picture and tends to find smoother edges.
When you should not upscale
If you can get a bigger original, do that instead. For print, check the print size you need at your printer's resolution; an upscaled picture is acceptable for a casual print but not a substitute for a real high-resolution photo. For a logo, vectorising it with PNG to SVG gives an image that scales to any size without any guessing, as long as the logo is simple.
A worked example: an old family photo scan
You have a small scan of a family photo and want a bigger copy for a frame. Start with Sharp resize at 2x and look at the result: edges are clean but the faces stay soft, as expected. Switch to AI with the Photo type at 2x and run it. The comparison slider now shows less blockiness in the sky and cleaner outlines on clothes, while the faces look more defined but not identical to the original. Because this is a family photo, keep the original scan next to the upscaled one, and use the upscaled version for display only.
FAQ
Is the first AI run slower?
Yes, because the model downloads. After that it works offline.
Why is my result not much sharper?
The source probably lacks detail, or you used Sharp resize on a very soft picture. Try the AI method with the right image type.
Does it add a watermark?
No.
Can I upscale a screenshot of text?
Yes with the Photo type, but verify the characters carefully; AI can alter small letterforms.