Rembrandt: a Local, Open Source Lightroom Alternative

A look at which AI photo-editing features Rembrandt really runs on your own hardware, how its self-hosted mode stores edits, and what to test before leaving Lightroom.

Binary Buddies · 5 min read · 8 views

Photo editors with AI masking, denoise and object removal usually come bundled with an account, a subscription and a cloud round-trip. Rembrandt, a project that surfaced on Show HN, makes the opposite claim: a free editor for macOS, Windows and Linux with no account, no tracking, RAW development, masks and AI features that all run on the device. The useful question for an engineer is which of those AI features can actually be local, and what that costs you.

This article works from the project's README, not from a test run. Where the README states a mechanism (a CPU inpainting network, GPU deconvolution, a 155 MB CLIP model), we repeat it. Where it gives no figure, such as speed or quality against Lightroom, we say so instead of guessing.

The project is also self-hostable, which changes the threat model for anyone who keeps a photo library on a NAS or a home server. That is where we start.

Self-hosting: one binary, XMP sidecars, no original touched

Besides the desktop builds, Rembrandt can be served from the machine that holds your photos and used from any browser. The install is a single line:

curl -fsSL https://raw.githubusercontent.com/thesnarkitecht/rembrandt/main/install.sh | bash -s -- --server

The README says the script asks which photo folder to use, starts the server at login and prints a private link. The server is described as one small, dependency-free binary, supported on Linux and macOS. Add --lan and other devices on your network can reach it.

The storage decision matters most. Edits are saved next to each photo as XMP, and the photos themselves are never changed. That means no proprietary catalogue database is holding your edit history hostage. If the project disappeared tomorrow, your originals are untouched and the edits sit in a documented sidecar format. It is also why a library on a NAS can be edited from a laptop without a sync layer in between.

Typing and talking to the editor, without a language model

The most surprising feature is natural-language editing. You press Ctrl/⌘ K and type 'warmer and a bit brighter', 'down exposure by ten points', 'shadows +25' or 'paste the edits from the previous photo'. There is also a microphone button in the side panel, so you can say 'bring lens blur down'. The sliders move while you watch.

The README is explicit about the implementation: it runs on your device and is 'a vocabulary, not a language model'. That is a sensible trade-off. A command grammar over a known set of sliders is cheap, deterministic and needs no multi-gigabyte download. The cost is that phrases outside the vocabulary will presumably not work the way they would with a real LLM. The README does not publish the vocabulary's coverage, so test the phrases you actually use before relying on it.

A related interaction needs no text at all: drag up or down on any part of the picture to lighten or darken that tone, left or right to change that colour, and the sliders follow.

GPU-heavy features and how they avoid hogging the machine

Several features are described as running on the GPU:

  • Refocus recovers detail in out-of-focus shots, 'even heavy defocus', using regularised deconvolution on the GPU. This is a classical signal-processing method, not a generative model, so it restores detail that was in the data instead of inventing it.
  • Super Resolution offers 2× and 4×, or Restore at the same size for soft photos. Metal is used on Apple silicon.
  • AI Denoise writes a new DNG that carries the same edits.
  • Merge builds HDR from brackets, panoramas and focus stacks, each to a RAW-like DNG.

The scheduling is the engineering detail worth noting. Denoise, Super Resolution and merges run in small GPU slices that pause while you edit, or when you are away, or overnight in a night window you set under Settings › Performance. The queue survives restarts. If you have queued a few hundred denoise jobs on a single laptop GPU, that design decides whether the machine stays usable. The README gives no throughput numbers, so how long a 4× upscale takes on your hardware is something you will need to measure.

Generative Remove on a CPU, and search without keywords

Generative Remove lets you paint over an object and fills the area with 'what was most likely behind it'. The README says this is a small inpainting network that runs on any computer's CPU, with no upload. A small model is the price of that portability, so expect it to be better on textured backgrounds and sky than on complex structure. 'People in the background' finds and removes passers-by in one click and keeps your subject. Heal and Clone remain for cases where copying from a matching part of the photo is more honest than generating pixels.

Smart search works the same way. Type 'dog on the beach' or 'red car' and it finds matching photos without keywording. It uses CLIP, a 155 MB model downloaded once when you turn the feature on in Preferences. The README says photos and searches never leave your computer. Note the one network event: the model download. If you are on an air-gapped box, you would need to check whether that download can be supplied another way, which the README does not say.

Learning your style from your own edits

Three features use your existing library as training signal, all on-device:

  • Edit like me learns from the photos you have edited, keyed on what each looked like before you touched it, so a dark photo still gets lifted the way you like. The README says there is nothing to train or upload.
  • Match my camera exploits the fact that every RAW carries the JPEG the camera made from it. Rembrandt learns your camera's colour and tone from those JPEGs, per camera model, and applies it with one switch.
  • Make these consistent matches exposure and white balance across a shoot to a hero frame you pick, by measuring the results, and keeps the rest of each photo's edits.

Culling gets similar treatment. Pick the best groups bursts, chooses the sharpest frame with open eyes in each, and rejects blurry and eyes-closed frames. Your own flags are kept, and there is a single Undo. That Undo is the safeguard worth checking first on a real shoot, because automatic rejection is exactly where a wrong model costs you a keeper.

Leaving Lightroom: what the importer brings across

Migration is where most alternatives lose people, and the README is specific. Presets import from .xmp, .lrtemplate and zip packs. Lightroom Classic catalogues come over with edits, keywords, labels, virtual copies and collections, 'with a report of anything that can't come over'. That report is the honest part: it means the project expects gaps and tells you about them. The README extract we have cuts off while listing further import sources, so we will not name more than Lightroom Classic here.

Day-to-day habits are covered too. Keyboard shortcuts can follow Lightroom's, and batch tools copy and paste edits to hundreds of photos. Watch a folder applies a preset and album to new photos as they arrive. Lens corrections use the camera's built-in profile from Fujifilm and Sony RAWs, or one of 1,500 Lensfun profiles, and RAW support is stated as 1,000+ cameras. Run the import on a copy of your catalogue first and read the report before trusting it.

What 'open source' and 'local' ask you to trust

The code is on GitHub, which lets you read what the editor and the install script do. Two practical points follow. First, the README says the builds are not code-signed yet. On macOS you go through System Settings → Privacy & Security → Open Anyway, and on Windows through More info → Run anyway. Second, the install is a curl | bash pipe, so read install.sh before running it on a machine that holds your library.

The mitigations on offer are checksums and an update path. All release files and checksums are published on the Releases page, and the in-browser Update button (Settings › About) installs the latest release, verifies it against the published SHA-256 sums and restarts the server. Also check the repository's LICENSE file yourself. The README text we read does not state the licence, and 'free' and 'open source' are not interchangeable if you plan to redistribute or build on it.

Where this does not fit

Rembrandt is a young project, and the README is the only evidence we have. Several things are unknown from it: how its RAW rendering compares with Adobe's on your camera, how fast the GPU features run on mid-range hardware, and how well the small CPU inpainting model handles hard cases. A professional with a deadline and a decade of Lightroom catalogues should pilot it on a copy, not switch over.

The self-hosted mode is supported on Linux and macOS only, so a Windows-only home server is not covered by the one-line install. It also trades a vendor for a maintainer: updates, bug fixes and camera support depend on one repository's activity. And if your workflow relies on cloud sync across devices or shared team catalogues, a local-first, sidecar-based design gives you files and folders to sync yourself, not a hosted service. For individuals who want their library, their edits and their AI on their own hardware, it is worth an afternoon of testing.

  • #Lightroom alternative
  • #On-Device AI
  • #Open Source
  • #RAW editor
  • #Rembrandt
  • #self-hosted photo editor

Comments

Add a comment
Sign in with Google to join the conversation