AI Image Upscaling

No. The model and inference engine are downloaded to your device, and upscaling runs entirely in your browser — not a single byte of your image leaves your machine. To verify: open your browser developer tools, switch to the Network panel, process an image, and you will see no upload requests.

About 3535 px on the long edge on desktop, and about 1200 px on mobile. Larger images are automatically scaled down to a workable size rather than rejected, so you still get a result. The limit comes from memory: the model always outputs 4× dimensions, so bigger inputs make the intermediate buffers grow quickly.

The first run downloads the inference engine (about 6–14 MB) and the matching model (2.4–17 MB), then compiles them into an executable graph. The browser caches these files, so from the second image onwards it is much faster. Switching to a different model (for example from anime to photo) downloads that model once.

Pick “Cartoon / illustration” for anime, game art and illustrations; pick “Photo” for photographs, screenshots and landscapes. Choosing wrong makes results noticeably blurrier — the anime model treats real texture as noise and smooths it away. It is the most common mistake; once you pick correctly, we remember your choice.

The model natively outputs 4×, so 4× needs a single inference pass. 8× and 16× run two 4× passes, which raises both time and memory a lot, so the maximum input size is smaller. 1× and 2× also run 4× first and then scale back down — the benefit is cleaner lines and less noise.

Input: JPG, PNG, WebP and other common formats. Output: JPG or PNG — JPG is far smaller and suits photos and screenshots, while PNG is lossless but much larger. If the source has a transparent background we always output PNG, because JPG has no alpha channel and would turn transparent areas black.