Upscayl
FreeEnlarge a small blurry image — detail reconstructed, not stretched
A desktop tool that enlarges images with AI super-resolution models. Rather than interpolating pixels, the model infers and reconstructs detail — useful for restoring old photos, enlarging low-resolution assets and rescuing small images from the web. It runs fully offline, handles batches, and adds no watermark. The Windows installer is about 249 MB.
Why we picked it
Online upscaling services follow a familiar script: a few free images, sign up for 2×, pay for 4×, watermarked exports — and you have to upload your photos to get any of it.
Upscayl brings the job back to your machine: 48,227 stars, AGPL-3.0, all computation local, no uploads, no image limits, no watermarks. The interface needs no tutorial — pick an image, pick a model, pick a scale, press go.
Set expectations about results. It uses Real-ESRGAN-class super-resolution models, which work by inferring what detail should exist at higher resolution and adding it. On photographs, illustrations and anime art the results are usually excellent, and blurry old photos gain noticeable clarity. On text and fine charts results are mediocre and details can be reconstructed incorrectly. Several models ship in the app (realistic, anime-oriented, conservative), and the same image can look quite different between them — try more than one.
Practical touches: batch processing across a whole folder, a before-and-after comparison view, and output as PNG, JPG or WebP.
Two things to state:
- A GPU is needed. Acceleration goes through Vulkan, covering NVIDIA, AMD and Intel integrated graphics, and Apple silicon works too. Pure CPU is painfully slow and not recommended.
- Releases are infrequent. The current stable build, v2.15.0, dates from December 2024, and while the repository still saw commits through 2026, stable updates are not frequent. What exists is complete and usable — just do not expect a stream of new features.
One caution: detail added by AI is inferred, not recovered from the original. For evidence, authentication or anything requiring faithful reproduction, that distinction matters.