LocalSR A little more in every pixel.
Image restoration, kept localv0.1.3-beta

Give your images
a little more room.

That photo you kept. The scan you found. The image that deserves a second look.

Enlarge it, soften the noise, or work on the blur. LocalSR is a free, open-source desktop app that runs the models on your own computer. Your photos and videos are never uploaded, and there is no account.

Download the beta macOS · Windows · Linux

01 On the light table

Marseille, France

Concrete.
Color.
A little detail.

4×Enlargement
Low-resolution photograph of colored balconies and concrete railings in Marseille. The same photograph enlarged four times by LocalSR’s SPAN NomosUni model.

A real run, with room to inspect. This photograph was cropped and reduced to 480 × 240 pixels, then enlarged to 1920 × 960 with LocalSR’s Quick Start model. No extra blur was added to the input.

Photo: Tim Ziegelbaum
Model: SPAN NomosUni, Philip Hofmann
How this example was made

02 At your own desk

Local by design

Your images.
Your machine.

A desktop workspace for a patient second look. Choose a model, check the result, keep what works.

A work print of the processed Marseille photograph.
Marseille / Study 01LocalSR · 4×
  1. Start with what you have.

    Open a photograph, screenshot, illustration, or scan. Work on a single image or build a batch queue.

    JPEG · PNG · WebP · TIFF · DNG
  2. Choose the treatment.

    Quick Start gives you a place to begin. Choose a different model for photo enlargement, illustration, denoising, or deblurring.

    Models download when you choose them
  3. Look closely. Then save.

    Inspect the preview and save a separate result. The processing happens here, on your computer. Your source image stays yours.

    No cloud processing · No account

03 Useful particulars

A few things
worth a closer look.

SR stands for super-resolution. It estimates a larger image from the pixels you already have. The result is a model’s interpretation, so the final judgment is yours.

Room to go larger.
Choose 2×, 3×, or 4× output dimensions. Large images are processed in overlapping tiles to work within available memory.
Restoration, at the same size.
Dedicated models can work on image noise, JPEG artifacts, or motion blur without enlarging the image.
Color deserves care.
ICC-aware image output and support for common photo formats, including DNG input. Metadata handling depends on the format and settings.
LabsVideo is an experiment.
Phone and camera videos can be upscaled frame by frame, with timing, rotation and audio kept. Optional SeedVR2 models and HDR preservation are there to explore. Video is slow on any computer; start with a short clip.

04 Get LocalSR

A desktop application

Take it
from here.

Choose the installer for your computer.
Bring an image you’d like to revisit.

BETASigned and notarized.
Updates itself from here.

v0.1.3-beta
Released

macOS

Apple Silicon · macOS 14+
Also available: v0.1.2-beta (415 MB) — see the download page

v0.1.3-beta · MPS backend · CPU fallback · 414 MB
Download .dmg

Windows

Windows 10 / 11 · x86-64

v0.1.3-beta · CPU backend · unsigned installer · 168 MB
Download .exe

Linux

Linux · x86-64 · glibc 2.39+ (Ubuntu 24.04+)

v0.1.3-beta · CPU backend · AppImage · 613 MB
Download .AppImage

The macOS installer is Developer ID signed and notarized by Apple, and future betas arrive as in-app updates; it is the most tested build. The Windows and Linux packages run on the CPU, are not signed and do not update themselves yet. Every package ships with its full source and license notices. LocalSR is provided as is under the MIT license, without warranty.

Before you begin

A few good questions.

Is it free?

Yes. LocalSR is open source under the MIT license, with nothing to buy or sign up for. The models you download keep their own licenses; the model guide lists them.

Do my images leave my computer?

No. Photos and videos are processed on your computer and never uploaded. You need an internet connection to download models; after that, they run locally. The app checks for updates at launch and sends an anonymous count with that check (app version, platform and channel, no identifiers), which you can turn off. The website counts visits without cookies. The privacy page has the details.

What does “beta” mean here?

Features can still change, and the Mac build is the most tested: it is signed, notarized and tells you when a newer beta is available. The Windows and Linux packages pass the same automated checks but have not had the manual testing the Mac build gets. Keep your originals. The release notes spell out the current limitations, and bugs go to GitHub Issues.

Will it use my graphics card?

On a Mac, yes: the app uses the Apple Silicon GPU through Metal (MPS), with a CPU fallback. The Windows and Linux packages run on the CPU for now; GPU editions for NVIDIA, AMD and DirectML are still held back, as the beta notes explain.

Can it recover the original detail?

Super-resolution estimates plausible detail from an image. It cannot guarantee the recovery of information that was lost, and results vary by source and model. Inspect small text, faces, and fine patterns carefully.

Are the models included?

Most models download separately, when selected. The optional HAT-S and HAT-L Face variants require a verified user-supplied checkpoint; read about their results and availability. LocalSR checks curated downloads against their expected size and SHA-256 checksum. Model sizes and licenses vary; some commercial-use terms still need clarification. See the model notes before using a particular model commercially.