AI Photo Upscaler
Enlarge blurry or small photos ×2 or ×4 with AI. No sign-up and no upload to a server: everything is processed right in this browser.
Loading the tool...
How to use
- Drag and drop one photo or click to choose it.
- Choose the upscale factor (×2 or ×4). ×2 is usually enough.
- The AI splits the photo into small pieces (tiles) and enlarges them one by one. A progress bar shows how far along it is, and you can cancel at any time.
- When it's done, compare the original and the result with the comparison slider.
- Download as PNG or JPG.
If the longest side of a photo is over 1500px, it is automatically scaled down before processing to protect memory. To keep the fine detail of a very large original, crop it to a suitable size before uploading.
What AI upscaling does
There are two ways to enlarge a photo: simply stretching the pixels, or having AI draw in new detail. The "resize" in Photoshop or a basic photo app only interpolates the existing pixels mathematically, so the bigger you go, the more smeared and blurry the photo looks. The AI upscaling (super-resolution) this tool uses is different: a neural network trained on huge numbers of photos infers that "this blurry pattern was most likely a line or texture like this" and fills in natural-looking detail. That's why an AI ×2 enlargement looks much crisper than an ordinary one.
Keep in mind that this inference doesn't "restore the original"; it "creates plausible new detail." AI-made detail can differ from the real thing, and parts where accuracy matters, like small text or fine facial features, can come out in shapes that don't match reality. It's great for social posts, prints or bringing old family photos back to life, but it must not be used for photos where "exactly as it was" matters, such as IDs, documents or legal evidence.
Which photos work best?
It works best on portrait, landscape and product photos that are slightly out of focus or taken at low resolution. If the blur isn't severe and the composition and colors are already good, the AI reinforces outlines and textures for a noticeably sharper result. Photos with heavy motion blur or that are extremely small (for example under 100px wide) give the AI too little information to work with, so they may not improve as much as you hope. The model was trained on real photos (people, landscapes, objects), so hand drawings, anime-style illustrations and scanned documents full of text can come out a bit odd, with smeared colors or thickened lines.
×2 vs. ×4: what's the difference?
×2 doubles the width and height, ×4 quadruples them. Internally the AI model always enlarges 4× first; when you choose ×2, the result is then precisely scaled down by half for a sharper, more stable 2× result. So ×2 is a quick, safe choice for general deblurring or social media, and ×4 is good when you need to blow up a small photo to print size. The number of tiles is the same for ×2 and ×4, so the amount of computation is similar, but the ×4 file is much larger and can take longer on slow devices.
Why process in tiles?
Running AI upscaling on a whole photo at once uses a lot of memory. Especially when it runs only in your own browser with no server, as this tool does, even a moderately large photo can run out of memory and freeze the browser or throw an error. So this tool cuts the photo into small tiles of around 192×192px, feeds them to the AI model one at a time, and stitches the results back together into the full photo. To avoid visible seams between tiles, each tile is cut with an overlapping margin, computed, and then the overlap is trimmed before stitching.
Thanks to this, you can upscale photos on an ordinary office or home PC without installing anything or needing a powerful graphics card. The bigger the photo, the more tiles there are and the longer it takes, so the progress bar shows how many tiles are done out of the total. Both ×2 and ×4 run the same 4× AI model (×2 is then scaled down by half), so the AI work is about the same; ×4 just produces a much larger file, which takes longer to save and display and uses more memory. If you want to switch photos or leave the page mid-way, press Cancel to stop at any time. Very large photos (over 1500px on the longest side) are automatically scaled down first to prevent out-of-memory failures.
Open-source software used
Real-ESRGAN (realesr-general-x4v3) by Xintao Wang et al., BSD 3-Clause License. A super-resolution AI model that enlarges photos 4×. The weights, converted to ONNX format, are hosted directly on this site (assets/vendor/models/realesr-general-x4v3/) and are never loaded from an outside server or CDN. The full license text is in LICENSE.txt in the same folder.
ONNX Runtime Web by Microsoft, MIT License. A library that runs AI models in the browser with WebAssembly. Its files are hosted directly on this site (assets/vendor/onnxruntime-web/). The full license text is in ort.LICENSE.txt in the same folder.
FAQ
Is my photo uploaded to a server?
No. The AI model and all processing run only in this browser. Your photo is never sent anywhere.
Should I choose ×2 or ×4?
For general deblurring or social media, ×2 is often enough and it's faster. Choose ×4 when you need to make a small photo very large, for example for printing. Note that ×4 takes longer to process.
Can I use it on ID photos or evidence?
We don't recommend it. AI upscaling infers and fills in detail that wasn't in the original, so it can create shapes (letters, facial outlines and so on) that differ from reality. Don't use it for IDs, documents or photos used as legal evidence.
Does it work on drawings or illustrations?
This tool uses a model optimized for real photos, so drawings, line art and comics may come out with smeared colors or blurred lines. Check the result carefully when you use it on illustrations.