LocalSR v0.4.0-beta

LocalSR v0.4.0-beta

Release candidate, not published yet (prepared 2026-11-10).

What changed

This is the first public release since v0.3.0-beta. It adds support for built-in
graphics and Intel NPUs, updates inside the app on Windows and Linux, comparing saved
results, HEIC and MOV output on the Mac, and a long pass over reliability and the
interface. If you run the Windows or Linux build of v0.3.0-beta, install this version
once from the download page; it uses the same settings, queue and models folder, and
later versions update in the app.

Since v0.3.0-beta

• Graphics and NPUs. On Windows, DirectML can use built-in graphics (Intel, AMD
  Radeon), and Intel graphics also appear as OpenVINO devices on Windows and Linux. A
  GPU is used only after LocalSR has compared its result with the CPU's, and a driver
  that crashes during that test is left alone. Intel NPUs run SPAN, Real-ESRGAN, FBCNN and
  NAFNet on Windows and Linux.
• Updates on every system. Windows and Linux builds now update themselves, with
  signed update statements that also cover package sizes and parts. The update switches
  in the LocalSR menu stop the check at launch or the anonymous count.
• Fixes before upscaling. Choose the model for each fix, run up to three in order
  (noise, JPEG, blur), and see in the Model plan how many models fit the task. A fix the
  model cannot process is skipped with a note instead of ending the job.
• Compare. Pick up to four results in Outputs and compare them with a slider or
  linked boxes at full resolution. "Try others" queues two to four installed models on
  one image. Videos use the same slider and linked boxes with synchronised playback
  from one transport.
• Output formats. On the Mac, HEIC images and MOV video with HEVC or Apple ProRes.
• Video. Optional frame interpolation (2x, 4x, 60 or 120 fps), a Labs feature that
  is off by default.
• Photos. Queued photos with the same plain recipe (SPAN, Real-ESRGAN or FBCNN, no
  fixes or Faces) run as a batch on Apple Silicon, about 1.9 to 2.8 times faster than one
  at a time on an M1 Pro. Grey scans keep their grey, and 16-bit grey and CMYK profiles
  are read correctly.
• Models. Two DAT2 models join the library, the Hardware menu groups each graphics
  card's routes, and the model library shows each author, source and licence.
• Packaging. Linux AppImages need glibc 2.35, so they run on Ubuntu 22.04, and the
  ROCm and CUDA ones are about a third smaller. On the Mac, Homebrew installs LocalSR
  and closing the window no longer quits the app.

Reliability

• H.264 and HEVC export accepts low-sample-rate audio without "Cannot Encode Media".
• With Video format Automatic, an early system HEVC encoder refusal can continue as AV1
  with a note.
• A result is never written over an existing file, including names that differ only in
  case or Unicode form.
• Before a job starts, LocalSR checks that its files fit on the output's disk and keeps
  8 GB free on the system volume. A full disk stops the queue once with a plain message.
• The remaining-time estimate is rewritten. It starts only once a steady speed is
  measured, counts down by itself and changes in steps that do not flicker.
• Core ML preparation has a 120 second limit per job. A program that does not finish is
  remembered and tried again after a week or after an app or macOS update. Intel NPU
  models are compiled in a separate process.
• Engine processes are stopped as a whole group, and a damaged queue database is set
  aside so the app starts.
• FFmpeg and the editing runtime are chosen only through the app's own file dialogs.
• Editing on a Mac no longer stops for 2 GiB of swap.

Interface

• Finished videos appear in Outputs with their duration and frame rate.
• Browse other inputs and results while a video job runs.
• Two start buttons: "Only this one" and "Upscale all N".
• Inputs and Outputs tabs with counts; Compare opens only from "Compare these N".
• A clearer running row, "Saving…" and "Stopping…" states, and removing inputs while a
  queue runs.
• Click a result to show it, "Use as input" to use it as a new input, Clear on Outputs.
• Help "?" tips on the Advanced controls and a hardware line on every library row.
• The Illustration choice picks illustration models: Real-ESRGAN Anime 6B for Quick
  and RealPLKSR HFA2k for Best.

Download

Download page. Checksums are in
SHA256SUMS and the .sha256 files.

Known limitations

• The Windows and Linux builds of this version were tested by automated checks: the
  installed app starts and its engine answers. The hands-on testing was on an Apple
  Silicon Mac.
• The Illustration choice was changed to the illustration models without a visual
  comparison against the previous behaviour.
• The CUDA editions have never run on an NVIDIA GPU, and DirectML's NAFNet denoising
  misses its accuracy tolerance. Intel NPUs on Windows have not been tried on real
  hardware, and AMD NPUs are not supported.
• Frame interpolation is a Labs feature; its 1080p and 4K speed and memory use on the Mac
  are not validated. De-flicker, editing, SeedVR2 and HDR preservation also remain
  experimental.
• Windows installers are not code-signed, so SmartScreen asks first.

See KNOWN_LIMITATIONS.md
for the current list. Report problems on
GitHub Issues with your app version and
steps to reproduce.

LocalSR v0.3.0-beta

LocalSR v0.3.0-beta

The Mac's Neural Engine does the work, and videos upscale about twice as fast.

What changed

The Neural Engine on Apple Silicon. Quick photo (SPAN), Real-ESRGAN and JPEG
cleanup (FBCNN) now run on the Neural Engine through Core ML, and Best photo
(RealPLKSR) runs on the GPU through Core ML. On a MacBook Pro M1 Pro that made a
tile two to three times as fast as before (SPAN 19 against 38 ms, Real-ESRGAN
anime 95 against 319 ms, FBCNN 61 against 149 ms) while using a fraction of the
energy. HAT, SwinIR, NAFNet and SCUNet stay on the GPU as before.

The first job with a model and tile size converts it into a Core ML program once
(1 to 40 seconds) and keeps it in ~/Library/Caches/LocalSR/coreml. LocalSR then
runs the first tile both ways and keeps Core ML only if it matches the usual
result and is not slower on your Mac; otherwise the job continues on the GPU.
That check is stored per Mac and macOS version, so newer Macs, whose GPUs can be
faster than their Neural Engine, choose for themselves.

Hardware: Neural Engine. The Neural Engine is its own entry under
Advanced › Hardware and the default on Macs that have it. Apple GPU (Metal) keeps
every model on the GPU. Before a job, the status bar names where the selected
model will run, or says that it is unsupported on the Neural Engine.

Faster video. Quick runs each frame of a video in one call instead of tiles,
and decoding, upscaling and encoding now overlap. A 15-second 720×1280 iPhone
clip upscaled 4× went from 1.6 to 3.6 frames per second: about two minutes
instead of five.

Automatic video format. The new default format hands frames straight to the
Mac's hardware HEVC encoder while the next ones are upscaled, without the
temporary lossless copy. No Mac encodes AV1 in hardware, so MKV exports, HDR
preservation and Windows and Linux use AV1. AV1, VP9, FFV1, H.264 and HEVC can
still be chosen directly; a previous AV1 default becomes Automatic once.

Play before upscaling. A video plays in the preview before it has a result,
with the same controls as a finished comparison. While it is being upscaled the
preview shows the live tiles, and the comparison follows when it finishes.

Zoom beside the playback bar. The zoom controls next to the video bar no
longer disappear and leave a gap when the pointer leaves the video.

Download

Download page. The Mac mini
serves the files through the Infomaniak VPS; the attached GitHub files are
mirrors.

• LocalSR-v0.3.0-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Developer ID signed and notarized by Apple. Installed macOS betas receive
  this version through the app's update check; the update includes the
  inference engine.
• LocalSR-v0.3.0-beta-Windows-CPU-x86_64.exe — Windows 10 or 11, x86-64.
  Unsigned, so SmartScreen asks first. DirectML and NVIDIA CUDA editions are
  attached as well and remain untested on those GPUs.
• LocalSR-v0.3.0-beta-Linux-CPU-x86_64.AppImage — x86-64, glibc 2.39 or
  later. AMD ROCm and NVIDIA CUDA editions come in parts.
• *-source-and-notices.tar.gz — source, build instructions and third-party
  notices for each build.

File checksums are in SHA256SUMS and the .sha256 files.

Validation

The Neural Engine path was measured on a 16 GB M1 Pro with the shipped
checkpoints, against PyTorch on the same pixels: SPAN 54 to 57 dB, Real-ESRGAN
71 dB, FBCNN 72 dB, RealPLKSR 69 to 70 dB, with PyTorch 2.13 and 2.14. Whole
images and videos were compared with the GPU path and an untiled reference, the
video speed on a real phone clip, and the packaged worker was frozen and run on
the Neural Engine. The Python (784), interface (191) and host (86) tests pass.
The Windows and Linux packages went through the same automated build and smoke
tests as v0.2.2-beta.

Known limitations

Only the M1 Pro has been measured; newer Macs decide between the Neural Engine
and the GPU on their first tile. Preparing a model uses up to about 4 GB of
memory once. De-flicker, editing, SeedVR2, HDR preservation and video face
processing remain experimental. See KNOWN_LIMITATIONS.md
for the current list.

Report problems on GitHub Issues
with your app version and steps to reproduce.

LocalSR v0.2.2-beta

LocalSR v0.2.2-beta

Videos are compared the way photos are, with playback controls below the
picture.

What changed

One stage for photos and videos. A finished video now sits on the same
canvas as a photo. The divider stops at the edges of the frame instead of
running across the empty canvas, and the zoom controls, now with 1:1, appear
while the pointer is over the video. Scrolling, pinching and View › Fit to
Window and Actual Pixels work for videos too. Original · Split · Enhanced and
holding M work as before.

Playback below the frame. Play and pause, the time, a thin scrubber and the
speed (0.5×, 1×, 1.5× and 2×) sit in one bar under the video, in the same style
as the zoom controls. Space plays and pauses. Both versions still play in
sync.

Rotated videos with PyAV 19. Videos whose track carries a rotation (a
display matrix, common on phones) stopped the engine when it ran with PyAV 19,
which changed how frame side data is read. The engine now reads it in a way
that works from PyAV 12 to 19. The published packages ship PyAV 18.1 and were
not affected; installs from source that picked up PyAV 19 were.

Download

Download page. The Mac mini
serves the files through the Infomaniak VPS; the attached GitHub files are
mirrors.

• LocalSR-v0.2.2-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Developer ID signed and notarized by Apple. Installed macOS betas receive
  this version through the app's update check; the update includes the
  inference engine.
• LocalSR-v0.2.2-beta-Windows-CPU-x86_64.exe — Windows 10 or 11, x86-64.
  Unsigned, so SmartScreen asks first. DirectML and NVIDIA CUDA editions are
  attached as well and remain untested on those GPUs.
• LocalSR-v0.2.2-beta-Linux-CPU-x86_64.AppImage — x86-64, glibc 2.39 or
  later. AMD ROCm and NVIDIA CUDA editions come in parts.
• *-source-and-notices.tar.gz — source, build instructions and third-party
  notices for each build.

File checksums are in SHA256SUMS and the .sha256 files.

Validation

The video comparison was checked in the interface tests (186 pass) and with the
real preview in a browser on a 4× result of a 320×240 MJPEG clip from 2007:
the video fits the canvas, the divider stays on the frame while dragging, 1:1
shows 228 %, and play, pause, scrubbing, speed and Space work. The same clip
was upscaled with Best in a development build of the desktop app on a 16 GB
M1 Pro.
The rotation fix passes the video timing and HDR tests with PyAV 18.1 and
19.0.2 and was also run against PyAV 12.3. The packaged Mac app, its engine and
every Windows and Linux package went through the same automated build checks
as v0.2.1-beta.

Known limitations

Editing, SeedVR2, HDR preservation, de-flicker and video face processing remain
experimental. See KNOWN_LIMITATIONS.md
for the current list.

Report problems on GitHub Issues
with your app version and steps to reproduce.

LocalSR v0.2.1-beta

LocalSR v0.2.1-beta

Comparison controls that stay out of the way, and a divider that stays put
when you zoom in.

What changed

The divider stays where it is. Zooming and panning used to move the
comparison divider with the picture, so a close look at a corner pushed it off
screen and showed only one side. It now keeps its place, its 2 px line and its
handle at every zoom; both images move underneath it, so at 1000 % you still see
the original on the left and the result on the right. Fit keeps the divider
where you put it.

Original · Split · Enhanced. The two labels above the preview are now a
quiet switch: Original shows the whole original, Split the divider, Enhanced the
whole result. Your choice stays until the next result, which starts in Split.
The same three views are in the View menu. Hold M to show the original from
any view; let go to return. Images and videos work the same way.

Zoom controls only when you need them. The zoom box moved from the top right
of the picture to the bottom right and appears while the pointer is over the
image. Scrolling, pinching and View › Fit to Window and Actual Pixels work as
before.

Download

Download page. The Mac mini
serves the files through the Infomaniak VPS; the attached GitHub files are
mirrors.

• LocalSR-v0.2.1-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Developer ID signed and notarized by Apple. Installed macOS betas receive
  this version through the app's update check; the update includes the
  inference engine.
• LocalSR-v0.2.1-beta-Windows-CPU-x86_64.exe — Windows 10 or 11, x86-64.
  Unsigned, so SmartScreen asks first. DirectML and NVIDIA CUDA editions are
  attached as well and remain untested on those GPUs.
• LocalSR-v0.2.1-beta-Linux-CPU-x86_64.AppImage — x86-64, glibc 2.39 or
  later. AMD ROCm and NVIDIA CUDA editions come in parts.
• *-source-and-notices.tar.gz — source, build instructions and third-party
  notices for each build.

File checksums are in SHA256SUMS and the .sha256 files.

Validation

The comparison changes were checked in the interface tests (184 pass) and in
the running app on a 16 GB M1 Pro: zooming from Fit to over 1000 % and panning
left the divider in the same screen position with the same line and handle,
holding M and the three views showed the expected image, and the zoom controls
appeared only over the picture. The packaged Mac app, its engine and every
Windows and Linux package went through the same automated build checks as
v0.2.0-beta.

Known limitations

Editing, SeedVR2, HDR preservation, de-flicker and video face processing remain
experimental. See KNOWN_LIMITATIONS.md
for the current list.

Report problems on GitHub Issues
with your app version and steps to reproduce.

LocalSR v0.2.0-beta

LocalSR v0.2.0-beta

A calmer, simpler window, redesigned for people who open LocalSR for the first
time.

Why

The first minutes with LocalSR asked too much. Two or three places offered to
download a model, presets were named after model codenames, a top bar repeated
the app name next to buttons that already lived in the menu bar, and blue,
green and amber text competed with the photo you were trying to judge. This
release keeps what LocalSR does and changes how it asks you to do it.

What changed

One way to start. Add a photo, pick Quick or Best, press the button at the
bottom right. If the model is not on your computer yet, the button says so
("Download 4 MB, then upscale") and does both; while it downloads, the same
button cancels. Per-model download buttons, the first-run download notice and
the license box with its own download button are gone. A model's license is a
one-line credit with links to its source and terms.

Plain choices. Quick and Best are described by what you get, and each card
names the download it still needs. Upscale, Restore and Edit form one switch
with a single caption, and Upscale handles videos too, so the separate Upscale
Video task is gone. Models that are ready carry no label. A line beside the
button says what the next job will make, for example
2268 × 4032 → 9072 × 16128 · 4× · Apple GPU (Metal).

No top bar. On wide windows the pane headings form the top of the window,
with the window buttons beside Media. The model library opens from
View › Models… (⇧⌘M) or the Change… link under Model, the benchmark from
View › Performance and Diagnostics…, and updates from LocalSR › Check for
Updates… (Help on Windows and Linux). Narrow windows keep a slim bar to switch
between panes.

A quieter look. The app ships IBM Plex Sans, so Mac, Windows and Linux show
the same typeface. Corners are square, panels are neutral grey so colours are
judged against a neutral surround, and colour appears in one place: the button
that starts work and its progress, in the terracotta of the logo. Until there
is something to start, the window has no colour at all.

Comparisons read left to right. The original is on the left of the divider
and the result on the right, for images and video, matching the labels.

Fixed

• New installations start on the GPU (Apple silicon, CUDA or ROCm). A fresh
  install used to default to the CPU, so a first upscale on a Mac ran several
  times slower than necessary until Hardware was changed under Advanced.
• Choosing Best no longer shows two download buttons for one file, with two
  different sizes.
• The window fits small screens. On a 1366 × 768 laptop or a 1280 × 800 display
  the bottom bar with the start button used to sit under the taskbar.
• Edit no longer offers to download a model on a computer where the edit cannot
  run.

Download

Download page. The Mac mini
serves the files through the Infomaniak VPS; the attached GitHub files are
mirrors.

• LocalSR-v0.2.0-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Developer ID signed and notarized by Apple. Installed macOS betas receive
  this version through the app's update check; the update includes the
  inference engine.
• LocalSR-v0.2.0-beta-Windows-CPU-x86_64.exe — Windows 10 or 11, x86-64.
  Unsigned, so SmartScreen asks first. DirectML and NVIDIA CUDA editions are
  attached as well and remain untested on those GPUs.
• LocalSR-v0.2.0-beta-Linux-CPU-x86_64.AppImage — x86-64, glibc 2.39 or
  later. AMD ROCm and NVIDIA CUDA editions come in parts.
• *-source-and-notices.tar.gz — source, build instructions and third-party
  notices for each build.

File checksums are in SHA256SUMS and the .sha256 files.

Validation

On a 16 GB M1 Pro with the packaged app: first launch in an empty profile,
Quick and Best downloads (including cancelling and resuming one), upscales of a
JPEG, a 2268 × 4032 HEIC (to 9072 × 16128) and a short H.264 clip with audio, a
JPEG-artifact fix before upscaling, Restore, a three-image batch, a FLUX.2
klein edit, the model library, the benchmark, Check for Updates and a narrow
window. The packaged worker passed the real-model acceptance run on the GPU.
The Windows CPU installer was installed and used in a Windows 11 virtual
machine (download, upscale, comparison, menus; the window now fits its
1280 × 800 screen). A Linux CPU AppImage from the same night passed its
headless smoke test and a real upscale on Fedora 44. Every Windows and Linux
package also passed the build's own install-and-handshake smoke test.
767 Python, 179 interface and 84 host tests pass.

Known limitations

Editing, SeedVR2, HDR preservation, de-flicker and video face processing remain
experimental. See KNOWN_LIMITATIONS.md
for the current list.

Report problems on GitHub Issues
with your app version and steps to reproduce.

LocalSR v0.1.7-beta

LocalSR v0.1.7-beta

Large upscales now appear in the comparison view after saving.

What changed

A 12 MP phone photo enlarged 4× becomes a 195 MP image. LocalSR could save it,
but the preview decoder then refused to open it because it exceeded Pillow's
pixel limit. That left the finished result hidden behind a “too large” error.

Completed images now use a separate decoder when needed, which sends a preview
of at most 2048 pixels on the longest side to the app. Your saved image keeps
its full resolution. The app also reuses the latest preview instead of decoding
the same result twice, and refreshes it when the file changes.

The README has clearer wording, a real before-and-after example and a new
screenshot of the model library.

Download

Download for macOS.
The Mac mini serves the files through the Infomaniak VPS. The attached GitHub
files are mirrors.

• LocalSR-v0.1.7-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Developer ID signed and notarized by Apple.
• LocalSR-v0.1.7-beta-source-and-notices.tar.gz — source, build instructions
  and third-party notices for this build.

Installed macOS betas receive this version through the app's update check.
The update includes the inference engine. File checksums are in SHA256SUMS.

Windows and Linux stay on
v0.1.3-beta.

Validation

Reproduced the old failure with a 16128 × 12096 PNG (195 million pixels), then
confirmed that it opens as a 2048 × 1536 preview with the fix. Regression tests
cover PNG, JPEG, TIFF and WebP, preview reuse, file replacement and decoder
failure. The source-image decoder keeps its existing pixel checks.

Known limitations

Editing, SeedVR2, HDR preservation, de-flicker and video face processing remain
experimental. See KNOWN_LIMITATIONS.md
for the current list.

Report problems on GitHub Issues
with your app version and steps to reproduce.

LocalSR v0.1.6-beta

LocalSR v0.1.6-beta

Edit now runs on 16 GB Macs, and LocalSR opens HEIC photos and camera RAW files.

Download

• LocalSR-v0.1.6-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Signed and notarized; uses the Metal (MPS) backend.
• LocalSR-v0.1.6-beta-source-and-notices.tar.gz — source, build
  instructions and third-party license notices for exactly that build.

Every file's SHA-256 is in SHA256SUMS and in the .sha256 beside it.

Windows and Linux stay on
v0.1.3-beta
for now.

Installed macOS betas update themselves: the app checks the feed at launch,
verifies the download's size, SHA-256 and signature, backs up settings, recipes
and the queue, then installs and restarts. This update also replaces the
inference engine, so it downloads about 760 MB.

What's new

• Edit works on 16 GB Macs. The memory guard misread how much memory a
  Mac has free: it counted only free and inactive pages and kept 4 GB in
  reserve, so it refused FLUX.2 klein 4B on a 16 GB MacBook Pro that macOS
  itself reported as 70% free. It now asks macOS, keeps 2 GB in reserve, and
  watches the editor's real memory use. On that Mac, with other apps open, an
  edit at 512 px takes about two minutes and does not swap.
• More photo formats. HEIC and HEIF from iPhones and cameras, and JPEG XL,
  open on a Mac through macOS's own decoders. AVIF, JPEG 2000 and GIF open too,
  and camera RAW files from Canon (CR2/CR3), Nikon, Sony, Fujifilm, Olympus,
  Panasonic, Pentax and others, not only DNG.
• Progress you can see. An edit names what it is doing (reading the photo,
  reading your instruction, step 2 of 4, finishing the image) with a moving
  progress bar on the photo and in the status bar.
• The right model by default. A Mac starts with the best editing model that
  leaves most of its memory to macOS and your apps: FLUX.2 klein Q4_0 on 16 GB,
  Q8_0 on 24–32 GB, Qwen Image Edit from 36 GB. A model you chose is kept.
• A calmer memory guard. It lets macOS compress memory and swap a little,
  and stops an edit only when the Mac is really short (critical pressure or
  heavy swapping), saying why.
• Fixed: installing a newer disk image over a copy that had updated itself
  kept the older inference engine.
• From v0.1.5-beta: the disk image opens as a drag-to-Applications window, and
  the dot in the icon is centred.

Known limitations

• Edit is still experimental. FLUX.2 klein 4B Q4_0 at 512 px is validated
  on a 16 GB M1 Pro; FLUX.2 klein Q8_0, the Qwen bundles and larger edit sizes
  have not run on real hardware yet.
• Edits are 512, 768 or 1024 px on the longest side, one image at a time.
• HDR preservation, SeedVR2 3B, de-flicker and video face processing are Labs
  features with unverified perceptual quality.
• The two HAT face companions still require a manual import.

The full list is in KNOWN_LIMITATIONS.md.

Feedback

Report problems on GitHub Issues
with the app version, model, device and steps to reproduce; Copy diagnostics
adds the hardware details without file names or media.

LocalSR v0.1.5-beta

LocalSR v0.1.5-beta

Installing LocalSR on a Mac is now a single drag.

Download

• LocalSR-v0.1.5-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Signed and notarized; uses the Metal (MPS) backend.
• LocalSR-v0.1.5-beta-source-and-notices.tar.gz — source, build
  instructions and third-party license notices for exactly that build.

Every file's SHA-256 is in SHA256SUMS and in the .sha256 beside it.

Windows and Linux stay on
v0.1.3-beta
for now.

Installed macOS betas update themselves: the app checks the feed at launch,
verifies the download's size, SHA-256 and signature, backs up settings, recipes
and the queue, then installs and restarts. This update also replaces the
inference engine, so it downloads about 760 MB.

What's new

• A proper install window. Opening the disk image shows one window with
  LocalSR on the left, your Applications folder on the right and an arrow
  between them: drag the icon across and you are done. The disk itself shows
  the LocalSR icon.
• A centred icon. The terracotta pixel in the app icon and the in-app logo
  sat slightly right of and below the middle of its crop marks; it is now
  exactly centred.

Nothing changed in upscaling, restoration or editing; models and settings carry
over unchanged.

Known limitations

• Edit is still experimental. Its memory profiles are estimates and no
  real edit has been validated on the 16 GB M1 Pro used for development yet.
  Close other apps before editing.
• Edits are 512, 768 or 1024 px on the longest side, one image at a time.
• HDR preservation, SeedVR2 3B, de-flicker and video face processing are Labs
  features with unverified perceptual quality.
• The two HAT face companions still require a manual import.

The full list is in KNOWN_LIMITATIONS.md.

Feedback

Report problems on GitHub Issues
with the app version, model, device and steps to reproduce; Copy diagnostics
adds the hardware details without file names or media.

LocalSR v0.1.4-beta

LocalSR v0.1.4-beta

LocalSR can now edit a photo from a written instruction, on your own Mac.

Download

• LocalSR-v0.1.4-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Signed and notarized; uses the Metal (MPS) backend.
• LocalSR-v0.1.4-beta-source-and-notices.tar.gz — source, build
  instructions and third-party license notices for exactly that build.

Every file's SHA-256 is in SHA256SUMS and in the .sha256 beside it.

Windows and Linux stay on
v0.1.3-beta
for now; image editing is not in those builds yet.

Installed macOS betas update themselves: the app checks the feed at launch,
verifies the download's size, SHA-256 and signature, backs up settings, recipes
and the queue, then installs and restarts. This update also replaces the inference engine, so it
downloads about 760 MB.

What's new

• Edit with a prompt. Choose Edit beside Upscale and Restore, select an
  image and write what should change ("make it winter", "remove the sign").
  The result is a new PNG in your output folder; the original stays untouched.
  Progress, cancel and the before/after comparison work as for every other job.
• Editing models in the library. A new Edit with a prompt group lists
  Qwen Image Edit 2511 in six sizes (24 GB of memory and up) and FLUX.2 klein
  4B in two sizes, the choice for 16 GB Macs. All are Apache-2.0, downloaded
  only when you choose them (5 to 26 GB), pinned to an exact upstream revision
  and checked with SHA-256 before every use. Each one says what it needs and
  whether it fits this Mac.
• A memory guard. Editing runs in its own process. LocalSR checks free
  memory before it starts, and a small supervisor stops the edit if the Mac
  runs short or starts to swap, instead of letting the system grind.

Known limitations

• Edit is experimental. Its memory profiles are estimates: on the 16 GB
  M1 Pro used for development, the guard stopped both test edits because other
  apps left too little memory free, so no real edit has been validated yet.
  Close other apps before editing.
• Edits are 512, 768 or 1024 px on the longest side, one image at a time.
• Qwen Image 2.1 is not offered: its license allows research and evaluation
  only.
• HDR preservation, SeedVR2 3B, de-flicker and video face processing are Labs
  features with unverified perceptual quality.
• The two HAT face companions still require a manual import.

The full list is in KNOWN_LIMITATIONS.md.

Feedback

Report problems on GitHub Issues
with the app version, model, device and steps to reproduce; Copy diagnostics
adds the hardware details without file names or media.

LocalSR v0.1.3-beta

LocalSR v0.1.3-beta

Open now opens the file you selected, the app has a user guide, and it is
simply called LocalSR.

Download

• LocalSR-v0.1.3-beta-macOS-arm64.dmg — Apple Silicon, macOS 14 or later.
  Signed and notarized; uses the Metal (MPS) backend.

• LocalSR-v0.1.3-beta-Windows-CPU-x86_64.exe — Windows 10 or later, 64-bit.
  CPU processing. The installer is not signed, so SmartScreen shows a
  warning: choose More info → Run anyway, after checking the SHA-256.

• LocalSR-v0.1.3-beta-Linux-CPU-x86_64.AppImage — glibc 2.39 or newer
  (Ubuntu 24.04 and later). CPU processing. chmod +x it and run it.

• GPU editions. None of these is code-signed, and only ROCm has run on
  its GPU family; the others are published untested:

  • LocalSR-v0.1.3-beta-Linux-ROCm-x86_64.AppImage — AMD Radeon supported by
    ROCm 7.2 with the amdgpu driver. Ran on a Radeon RX 9060 XT in earlier
    builds. About 10 GB in five .part-000N files.
  • LocalSR-v0.1.3-beta-Linux-CUDA-x86_64.AppImage — NVIDIA GPU with a CUDA
    12.6 driver. Untested: it has never run on an NVIDIA GPU. About 7.6 GB
    in four parts.
  • LocalSR-v0.1.3-beta-Windows-CUDA-x86_64.exe — NVIDIA GPU with a CUDA
    12.6 driver. Untested. A 5 MB installer plus two
    .engine.tar.gz.part-000N files (2.7 GB); keep them in one folder and run
    the installer, which checks and unpacks the engine.
  • LocalSR-v0.1.3-beta-Windows-DirectML-x86_64.exe — any DirectX 12 GPU.
    Untested, and its photo denoising still differs measurably from the
    CPU's.

  GitHub stores files of up to 2 GB, so the AppImages come in parts. Join them
  with cat LocalSR-…AppImage.part-* > LocalSR-…AppImage, check the result
  against its .sha256, then chmod +x it.

• …-source-and-notices.tar.gz — one per package: source, build instructions
  and third-party license notices for exactly that build.

Every file's SHA-256 is in SHA256SUMS and in the .sha256 beside it.

Installed macOS betas update themselves: the app checks the feed at launch,
verifies the download's size, SHA-256 and signature, backs up settings, recipes
and the queue, then installs and restarts. This update also replaces the
inference engine, so it downloads about 730 MB. The Windows and Linux builds
have no updater yet — come back to the
download page for the next one.

What's new

• A user guide. Help → LocalSR User Guide and the app menu open a short
  guide to the three tasks, models, batches, video and where results are saved:
  https://herrei.github.io/localsr/guide/.
• A welcome on first launch. It points to the guide and says plainly that
  the update check sends an anonymous count (app version, platform, channel)
  and where to turn it off. It appears once.
• Windows and Linux. LocalSR now has a Windows installer and a Linux
  AppImage, both built from the same licensing-clean media runtime as the Mac
  build, each with its own corresponding source, plus GPU editions for AMD
  ROCm, NVIDIA CUDA and DirectML.
• The app is called LocalSR. The installer, menu bar, About box and Finder
  Quick Action no longer say "LocalSR Next Preview". Settings, models and the
  update channel are unchanged. An installed beta keeps its current file name
  when it updates; rename it in Finder if you like.

Fixed

• Open and Reveal used the wrong file. After a batch, the buttons opened
  whichever result finished last, whatever was selected. They now act on the
  selected item's result, including results from an earlier session.
• The comparison handle drifted with the image. When zoomed or panned, the
  ↔ handle followed the image's centre and grew or shrank with the zoom. It now
  stays in the middle of the preview frame at the same size.
• The app declared macOS 12 as its minimum, but its PyTorch libraries need
  macOS 14. Older systems now get a clear message from Finder instead of an
  engine that does not start.

Known limitations

• The CUDA and DirectML editions are untested: CUDA has never run on an
  NVIDIA GPU, and DirectML's NAFNet denoising misses the accuracy tolerance.
  Their installed-package tests ran on machines without a GPU. The standard
  Windows and Linux downloads process on the CPU, which is several times
  slower than a GPU. Intel XPU is not published (docs/licensing-media.md).
• The Windows installers are not Authenticode-signed and neither Windows nor
  Linux can update itself yet. On Windows and Linux, H.264 and HEVC need an
  FFmpeg you install yourself.
• H.264/HEVC exports through the system codecs go into MP4 (or MOV); choose
  MP4, or use an installed FFmpeg for MKV.
• HDR preservation, SeedVR2 3B, de-flicker and video face processing are Labs
  features with unverified perceptual quality.
• The two HAT face companions still require a manual import.

The full list is in KNOWN_LIMITATIONS.md.

Feedback

Report problems on GitHub Issues
with the app version, model, device and steps to reproduce; Copy diagnostics
adds the hardware details without file names or media.