Upscayl: Free Open Source AI Image Upscaler That Enlarges Photos Without Losing Quality

Upscayl is a free open source AI image upscaler that makes low resolution photos bigger without turning them into a blurry mess. You give it a small or old photo, pick a model, and it enlarges the image by 2x, 3x, 4x, or more, adding detail as it goes, so a tiny screenshot or an old family photo becomes something you can actually print or post. It has 52,236 stars, it is written in TypeScript, it is licensed under AGPL-3.0, and it works on Linux, Mac, and Windows. The part that sets it apart from the dozens of upscaler websites is where the work happens. Upscayl runs on your own computer, on your own GPU, and your images never leave your machine. There is no upload, no account, and no queue. The project collects no data from you at all.

Upscayl is built on Real-ESRGAN, an AI model that was trained to enlarge images and fill in the detail a simple resize would lose, and it runs that model with NCNN Vulkan, a fast inference framework that uses your graphics card. In plain English, the AI has seen millions of photos and learned what a sharp edge, a strand of hair, or a brick wall looks like, so when you enlarge a photo it can guess the missing detail instead of just stretching the pixels. That is why the results look like a real photo and not like a zoomed in thumbnail. The default model is Real-ESRGAN, and the app also ships models for digital art, like Ultrasharp, and for other styles, like REMACRI and Ultramix. You pick the model that matches your image, and if you know what you are doing, you can add your own custom models too.

Upscayl is more than a one photo at a time tool. You can upscale a whole folder of images in one go with batch upscaling, which is the feature that makes it useful for real work, like restoring a folder of old photos or preparing a set of product images. You can save images as PNG, JPG, or WEBP, upscale an image twice with the double upscayl option when you need it even bigger, and compare the original and the result side by side before you save. There is also a command line tool, upscayl-ncnn, for people who want to script it, and it works with ImageMagick, so you can build it into a batch workflow. If you upscale a photo to print it, a Pixels to Inches converter tells you the print size at your chosen DPI, and if you are putting the upscaled image on a website, an Em to Px Converter helps you size it in the stylesheet.

Upscayl needs a Vulkan compatible graphics card, and this is the one requirement that catches people out. Most modern dedicated GPUs work, and you can pick which GPU to use in the settings. Most modern integrated GPUs do not work, according to the README, so if your laptop only has integrated graphics, check first. The app is built on Electron, so it runs on Linux, Mac, and Windows, and the releases page has installers for all three, plus a portable version for Windows. There is no paid tier and no limit on how many images you upscale, because the work happens on your hardware. If your GPU is not supported, the app will tell you, and the community forum is the place to ask which cards people have had success with.

Download Upscayl from the releases page or from upscayl.org and install it. Open the app, choose your image or a whole folder, and pick a model. Real-ESRGAN is the default and a good start for photos. Ultrasharp is the one to try for digital art. Choose the scale, 2x, 3x, or 4x, choose the output format, PNG, JPG, or WEBP, and press the button. The app shows a progress bar and a side by side comparison when it finishes, so you can see exactly what the AI added before you save. If the result looks too smooth or too sharp, try a different model. That is the whole workflow, and for a folder of photos, the batch mode does them all while you do something else.

Upscayl is an upscaler, not a miracle worker, and the README is honest about the limit. It cannot de-blur an out of focus image. If the photo was blurry when you took it, enlarging it will give you a bigger blurry photo, with the AI guessing at detail that was never there. It works best on images that are sharp but small: old photos, screenshots, compressed images, and scans. It also cannot invent what is not in the image, so a tiny face in a crowd will not become a portrait, no matter how good the model is. Used for what it is for, though, it is one of the best free tools of its kind, which is why it has 52,236 stars and a community that keeps adding models and guides for it.