Free, open-source AI image upscaler that runs fully offline
Upscayl is a free, open-source desktop app that enlarges images with AI models from the Real-ESRGAN family. Pick an image, pick a model, pick a scale, and the job runs locally: no account, no upload, no watermark. It is available for Windows, macOS and Linux under the AGPL-3.0 license; the current stable release is version 2.15.0, published on December 25, 2024.
Where most 'free' online upscalers add watermarks, queues or sign-up forms, Upscayl stays on your machine, and a batch mode chews through whole folders at once. The project has passed 50,000 GitHub stars. The honest catch is hardware: the developers ask for a Vulkan-compatible GPU, and warn that most integrated graphics will not do.
In our hands-on test we drove the command-line engine of the official v2.15.0 build on a Linux machine with no GPU (software Vulkan standing in): a 160x100 JPEG came back as a 640x400 PNG in about a minute. Large text and shapes turned impressively crisp; a barely legible caption line was 'reconstructed' into confident gibberish, the AI inventing letters it could not read. We did not test the desktop interface in that setup. Verdict: an easy recommendation with a Vulkan-capable GPU, slow going without one.
Featured in our discovery: Upscayl: Give Your Images a Second Life
Real-ESRGAN family models enlarge images 2x, 4x, 8x or 16x; sharpening and detail recovery happen on your own hardware.
Point it at a folder and every image inside gets upscaled in one run.
Ultramix Balanced, Ultrasharp, Remacri, Digital Art and High Fidelity ship with the app, no extra downloads needed.
No account, no upload, no watermark; we verified in our test that the engine runs with the network off.
Yes. The desktop app is free and open source under the AGPL-3.0 license. The team also runs a separate paid cloud service, but nothing in the desktop app requires it.
The official FAQ requires a Vulkan-compatible GPU and warns most integrated graphics do not qualify. Our test proved the engine can run on software Vulkan, but slowly: about a minute for a tiny 160x100 image.