Upscale and restore photos and video on your own computer.
LocalSR is a free desktop app for AI image and video enhancement. It runs the models locally, shows each tile as it finishes and lets you compare the result with the original while the queue keeps working. Your media never leaves your machine.
Download · Website · Documentation · Models · Changelog
The beta on an M1 Pro. Photo: NASA, Apollo 17 (capture details).
- Upscale photos, illustrations, screenshots and anime by 2×, 3× or 4×.
- Restore before upscaling: remove noise, motion blur and JPEG artifacts, with an optional face-aware pass.
- Video. Every image model runs frame by frame with the source timing, rotation and audio preserved. Phone and camera recordings open directly on macOS. SeedVR2 3B adds temporal consistency, and HLG/PQ HDR can be kept instead of tone-mapped (both Labs).
- Say what you want, not which network. Pick Quick or Best, photo or illustration, and what to fix first. The plan card shows which model runs at each stage, its license, its size and whether it fits your hardware. Save the setup as a recipe.
- A model library with provenance. Checkpoints download on demand, mostly
from their authors' releases and otherwise from pinned mirrors checked against
the official files. Each is pinned by size and SHA-256 and shown with its
license and source (provenance). Your own
.safetensorsmodels work too. - A queue that keeps you informed. Images, videos and whole folders; real tiles and frame progress; estimates measured from completed work; side-by-side comparison, including synchronised video playback.
- Automation and benchmarks. A
localsrcommand for scripts and watch folders, and separate CPU and GPU benchmarks.
Inputs: JPEG, PNG, WebP, TIFF and DNG camera RAW; MP4/MOV, MKV/WebM, AVI, MPEG/VOB, transport streams, WMV, FLV, 3GP and OGV. Outputs: PNG, JPEG, TIFF and WebP; AV1, VP9 or lossless FFV1 video, and H.264/HEVC through macOS's own codecs or an FFmpeg you install. Metadata and colour · Video
v0.1.3-beta is on the download page:
- macOS, Apple Silicon, macOS 14 or later: a DMG (414 MB), signed and notarized. The app checks for updates at launch and verifies each update's signature.
- Windows 10 and 11, x86-64: a CPU installer, plus DirectML and NVIDIA CUDA editions that are untested on those GPUs. None is code-signed.
- Linux x86-64: a CPU AppImage, an AMD ROCm edition (tested on a Radeon RX 9060 XT) and an NVIDIA CUDA edition (untested).
The CUDA and ROCm editions are larger than GitHub's 2 GB file limit and come in parts. Intel XPU packages are held back. Platforms explains what has been tested where.
No model weights are bundled. The Quick model (4.5 MB) downloads on first use, the others when you choose them.
| Platform | Engines | Tested on |
|---|---|---|
| macOS 14+, Apple Silicon | MPS | M1 Pro: images, video, cancellation, recovery, updates |
| Windows 10/11, x86-64 | CPU, DirectML, CUDA | Windows 11 with an Intel UHD 620 (CPU and DirectML); CUDA untested |
| Linux, x86-64 | CPU, AMD ROCm, NVIDIA CUDA | Fedora 44 with a Radeon RX 9060 XT (CPU and ROCm); CUDA untested |
Start with a small image or a short clip. Large images and video can take hours, and 16 GB of GPU memory does not guarantee that a SeedVR2 job fits. Known limitations
The catalog holds 18 verified checkpoints: SPAN, RealPLKSR, HAT-S, HAT-L, Real-ESRGAN and SwinIR for upscaling; NAFNet, SCUNet, FBCNN and RealPLKSR for noise, blur and JPEG repair; SeedVR2 3B (FP16 and FP8) for video. Quick uses SPAN NomosUni and Best uses RealPLKSR NomosWebPhoto, both by Philip Hofmann under CC BY 4.0. The two HAT face companions have to be imported by hand because their checkpoint rights are unresolved. Model guide · Licenses
There is no account, and the media you process never leaves your computer. LocalSR makes two kinds of network request: the update check at launch and the model downloads you start. The update check also sends a count — the version, platform and channel, with no identifier — so the project can tell how many installs are active; Send an anonymous update-check count in the Software Update dialog turns it off. Copy diagnostics includes versions, hardware and memory figures, never file names or media. The full description is on the privacy page.
Svelte interface
↕ typed Tauri commands and events
Rust host · queue, settings, downloads, updates, file access
↕ JSON Lines over stdin/stdout
Python worker · PyTorch, Spandrel, PyAV, LibRaw, face detection, SeedVR2
The webview has no shell, filesystem or network API. The Rust host grants the worker access to the files you chose, keeps the queue in SQLite and verifies every download. Inference runs in its own process, so a GPU or memory failure ends the job rather than the app, and the host restarts the worker. Desktop guide · Worker protocol · Architecture decisions
You need Python 3.11, Node.js, a stable Rust toolchain and Tauri's platform prerequisites.
./local-ci.sh setup # development dependencies
cd desktop && npm run tauri -- dev./local-ci.sh runs every check that works on a development machine. Packaging,
signing and the release pipeline are described in Releasing.
The localsr command line works without the desktop app:
Automation.
Bugs and feature requests go to GitHub Issues. Include the version, model, device and steps; Copy diagnostics supplies the hardware details. Security reports: SECURITY.md. Anything else: hermes.reisner@gmail.com.
The application is MIT licensed. Models keep their authors' licenses, shown in the app before download. Bundled libraries and their terms are listed in THIRD_PARTY_NOTICES.md; every release ships the corresponding source for its LGPL components.