Image Upscaler
Make a small or blurry image bigger and sharper with AI, right in your browser.
Loading Image Upscaler… If nothing happens, please enable JavaScript.
When you enlarge a picture the ordinary way, the program has nothing new to work with: it spreads the existing pixels over a bigger area and smooths the gaps, so a 300 pixel thumbnail blown up to 1200 pixels looks soft, with fuzzy edges and blocky JPEG squares. An AI upscaler works differently. It is a neural network trained on millions of pairs of small and large images, so it has learned what sharp edges, hair, text, brickwork or leaves usually look like, and it redraws the enlarged image with plausible detail instead of blur. This tool uses Swin2SR, an open super resolution model, and runs it entirely in your browser. Your images are never uploaded, there is no account, no watermark and no daily limit.
Frequently asked questions
Are my images uploaded to a server?
What is the difference between Photo x4 and Fast x2?
Can AI upscaling recover details that were never in the photo?
Why is there a maximum image size?
Why is it slow on my computer?
What happens to transparent PNG images?
Which format should I download?
Do I have to download the model every time?
About Image Upscaler
There are two modes. Photo x4 makes the image four times wider and four times taller (sixteen times the pixels) with a model trained on real-world photos, which also cleans up JPEG artefacts, noise and mild blur, so it is the right choice for old phone pictures, screenshots of images saved many times, or product photos that came out too small. Fast x2 doubles the size with a much lighter model: it is quicker and suits images that are already clean, such as illustrations, logos and recent photos that only need a bit more resolution. The picture is processed in small overlapping tiles that are blended together, which keeps memory use low enough for phones and laptops and lets you follow the progress and cancel at any point.
Speed depends a lot on your device. Browsers with WebGPU use the graphics card and finish a typical image in seconds; without it the model runs on the processor, where a 256 pixel image takes about 15 seconds in Fast x2 and about a minute in Photo x4. For that reason each mode has a maximum input size: without a graphics card 512 pixels per side for x4 and 1024 for x2, and with WebGPU 1024 and 2048, which already produces images of up to 4096 pixels. After processing, drag the slider to compare the AI result with a normal enlargement, switch to actual size to inspect details, and download as PNG, WebP or JPEG, one by one or all together in a zip. The model is downloaded once from this site and your browser keeps it for next time.
Enhance! Television's favourite impossible trick
For decades, crime dramas have shown investigators leaning over a monitor and saying "enhance" until a blurry reflection in a pair of sunglasses turns into a readable licence plate. Engineers laughed at it, because classic image processing cannot create information that the sensor never captured: sharpening filters only exaggerate edges that are already there.
Deep learning changed the conversation, though not in the way television imagined. In 2014 the SRCNN paper showed that a small convolutional network could learn to upscale images better than bicubic interpolation, and later models trained with adversarial losses (such as ESRGAN in 2018) began producing convincing hair, grass and fabric. The catch is that these details are guessed from training data, not recovered, so two different models can draw two different faces from the same few pixels.
Swin2SR, the model used here, was published in 2022 and builds on the Swin Transformer, which looks at the image through small shifting windows. Its real-world variant was trained on photos that were deliberately damaged with blur, noise and heavy JPEG compression, which is why it is good at cleaning up images that have been saved and shared many times.