Any user interested in exploring AI image generation won't take long to notice that the final size leaves much to be desired. Free services often set a cap of 1024×1024, and from there you have three options: pay for a premium subscription, generate images locally (and burn through VRAM in the process), or bet on upscaling. The open-source app Upscayl enables that third path with a very intuitive interface and the ability to install additional models.
Image generation services have moved past the 'me too' phase and now focus on advanced features like prompt optimization and video processing. However, one aspect that often remains on the other side of the paywall is resolutions. Their free modes tend to enforce a 1024×1024 limit, and the best alternative is to do upscaling ourselves. AI also plays a strong role in that field, and one of the most complete options is Upscayl, which offers a free and open-source client.
How to Upscale Images Without Losing Quality with Upscayl
Its developer warns that Upscayl's main requirement is a graphics card or integrated chip with Vulkan support. That doesn't mean the program will stop working entirely, but its upscaling will take much longer on the CPU. Personally, I found no compatibility issues, and regardless of the chosen model, upscaling only took a few seconds.
Running Upscayl boils down to five steps: select the image, choose the model from the list, set the upscaling level (4x is the most common), specify the destination path (otherwise it will save a copy in the source folder), and finally confirm with a click on the Upscayl button. The settings section includes a logs window where you can observe errors and other details, theme adjustments in the interface (the list is quite extensive), options to save images in PNG or JPG, and compression and scale parameters (I recommend keeping the defaults).
Now, the most interesting aspect of Upscayl is that it supports secondary models. One of the packages is available on GitHub for free, and the program's wiki explains how to convert preexisting models, but this is more than enough to experiment. As always, some images work better with a specific type of model (for example, Ultrasharp was designed for photos, but does a good job with certain illustrations), and the idea is to try each one, taking advantage of the lack of restrictions. Link below!
Official site and download: Click here
How to download programs from GitHub: Click here
