INGEST
Import XMP folders or Lightroom catalogs as the source material for a style pack.
[ Manifesto ]
A RAW file is not a photo. It is a person. Faces of the people who chose you as their photographer, GPS of the places you frequent, the hours you work, the contexts of the scenes you tell, the editing habits you have built over years. That is what leaves your machine when you upload a RAW to a cloud editing service.
That material lives on servers you cannot audit, often in jurisdictions you never consented to, often next to value chains (proprietary model training, biometric indexing, surveillance datasets) whose terms you never signed.
RedLnx exists because there is another way.
Local. Auditable. Free. Forever.
[ What RedLnx Actually Does ]
RedLnx is a local Rust + ONNX desktop app for photographers. It ingests real edits, filters obvious bad sources, adapts to your style, and writes local XMP outputs without a cloud handoff.
FILTER
Skip neutral sources, duplicate or derived sources, and Adobe AI / Adaptive Color poison markers before training.
PROFILE
Build a Local AI style model from real edits you already trust, instead of starting from a generic preset.
OUTPUT
Write local XMP sidecars for one style or multiple styles in dedicated output folders.
[ How It Works ]
The pipeline is concrete: train from real edits, inspect the develop deltas, then process a batch with one look or several.
Train
Import XMP folders or Lightroom catalogs. RedLnx builds a local style profile from your real edits and filters obvious poisoned or neutral sources before training.
Preview
Inspect before/after slider deltas on sample photos, compare style behavior, and choose a representative cover image for the style pack.
Batch
Write local XMP sidecars for one style or multiple styles, fully offline, with dedicated output folders per style.
[ The Liberation Counter ]
How much you got back.
A personal screen inside the app. Counts the hours of life RedLnx gave back compared with editing by hand. Counts the money you did not pay to incumbents. And, beneath everything, four zeros that name what did not leave your machine.
0 Files uploaded
0 Faces indexed
0 Coordinates shared
0 Servers contacted
Faces, coordinates, timestamps, contexts: none of this left this machine. The
number lives in user_profile.json on your disk and is never transmitted.
[ Screenshots ]
Four real interface views. Console, Train, Post-Produce, Style Pack.
[ How RedLnx Runs ]
CAPABILITY STACK IMPACT
Native Desktop Runtime Rust desktop app with local storage and predictable offline execution. Training, preview, and batch output stay on your own machine.
Local AI Inference Local ONNX inference with automatic fallback: CUDA, then DirectML on Windows, Metal on macOS, WebGPU / Vulkan on Linux, then CPU. Acceleration uses what is available without turning the workflow into a cloud dependency.
Lightroom-Oriented Pipeline Train from edits, preview the develop deltas, and write local XMP sidecars. RedLnx automates develop settings without pretending to replace Lightroom's renderer.
No Hidden Network Path No telemetry, no cloud queue, no photo upload path, no account. Your catalog, faces, GPS, edit history stay local. Always.
[ LOCAL / PRIVACY / LICENSING ]
No subscriptions. No cloud queue. No hidden upload path.
RedLnx is a native Rust desktop app with local storage and predictable offline execution. Once installed, training, inference, preview, and batch output stay on your hardware.
The workflow is Lightroom-oriented and XMP-based. RedLnx automates develop settings locally; Lightroom still handles final rendering.
The project is AGPLv3, open source, and free to download, with no telemetry, no tracker scripts, no subscriptions, no locked feature tiers, no account.
AGPLv3 is the trap. A corporation that wraps RedLnx into a proprietary cloud SaaS has to release the entire stack under the same terms. The license is the doctrine, written in legal code.
[ Roadmap ]
The plan is public. Status icons: ✓ shipped, ⟳ in progress, ○ planned. Priorities shift with community feedback.
Phase 1 · Core Intelligence
v0.2- Full mask integration · graduated, radial, brush
- Refined culling, white balance, straightening
- Local AI denoise engine
Phase 2 · Experience & Reach
v0.3- Redesigned interface · keyboard-driven batches
- Multi-language support · 12 locales
- Style community hub · share trained styles, keep catalog private
Phase 3 · Universal Workflow
v0.4- Capture One catalog bridging
- DaVinci Resolve Photo integration · styles as 3D LUTs
- RawTherapee, darktable, open-source RAW processor support
Phase 4 · Sovereignty
v1.0- Plugin and extension API · community-built scoring, exports, bridges
- In-software RAW demosaicing and processing · sensor to file, transparent
[ Future vision · post v1.0 ]
Clustered local inference across a photographer's own LAN (no central server, peer machines share load). Federated training on free and open datasets (Wikimedia Commons, public-domain catalogues), never on user-uploaded RAWs, with differential-privacy weight aggregation. Sovereignty extended to community-grown, without compromising the original promise.
[ Honest Funding ]
Where every euro goes.
RedLnx stays free forever. No subscription, no paywall, no locked tier. If you want to help fund the next twelve months of operating costs and the remaining release work, a Kickstarter is coming in a couple of weeks. Until then, Ko-fi is open.
LINE ITEM COST
Windows code-signing certificate ~€95 / year
macOS notarization · Apple Developer Program ~€90 / year
redlnx.dev hosting and domain ~€100 / year
Hardware, GPU test matrix, electricity to keep it running ~€2,200 / year
Development hours · solo, four a day fixed, after the real job variable
Where the money never goes: paid marketing, growth hacking, paid acquisition. Centralized inference servers. Account systems. Telemetry. Analytics. A "founder" badge. A "premium" tier. Any of it.
[ Hall of Fame ]
[ Live in ~2 weeks · post-Kickstarter ]
When the pre-launch Kickstarter closes, the names of supporters who opt in will appear here. Single mono list, alphabetical, no tiers, no badges, no perks. Recognition only. Pseudonyms welcome.
- Names are opt-in: pledging never auto-publishes a name.
- Order is alphabetical, never by pledge amount.
- No platinum / gold / silver. No "Founders Circle". A name is a name.
- Pseudonyms, first-name-only, "Anonymous" all accepted.
- Every feature stays free for everyone, forever. Recognition is the reward.
[ FAQ ]
What does RedLnx output?
RedLnx writes local XMP sidecars for Lightroom-oriented workflows. It automates develop settings; it is not a final raw renderer by itself.
What can I train on?
You can train from XMP folders or Lightroom catalogs. RedLnx profiles your existing edits and builds a Local AI adaptive style model.
Does RedLnx filter bad training data?
Yes. RedLnx already skips neutral sources, duplicate or derived sources, and Adobe AI / Adaptive Color poison markers before training.
Can I process the same batch with multiple looks?
Yes. RedLnx can write separate outputs for multiple styles in dedicated folders.
Do I need a GPU?
No. CPU works too. On Windows, RedLnx can use CUDA when the NVIDIA runtime is ready, otherwise DirectML or CPU fallback. On macOS, Metal. On Linux, WebGPU / Vulkan. GPU acceleration helps some parts of the pipeline more than others.
Does it work offline?
Yes. Training, inference, preview, and batch output are local once the app is installed. There is no cloud queue, no upload path, no account.
What affects training quality?
Training quality depends on the quality and consistency of the source edits. Cleaner source work produces a stronger style model.
Is there a subscription or paywall?
No. RedLnx is AGPLv3, open source, and free to download. There is no subscription tier, no locked feature, no premium plan. Every feature is for everyone, forever.
How was RedLnx built?
With AI-assisted development tools, openly. The idea sat in my head for years: a local, sovereign editing assistant that simply did not exist. Working four hours a day after the real job, alone, I would never have shipped it without those tools. Every line is reviewed, tested, and decided by me; the architecture, the AGPL choice, the doctrine, and the manifesto are mine. AI tools are instruments, not authors, the way a camera is not the photographer. The codebase is AGPL-3.0, fully open and auditable: if anything does not match what is written here, open an issue. The shipped app runs 100% offline, with no telemetry and no cloud calls; AI helped write the code, it does not run inside it.
What is the Liberation Counter?
A personal screen inside the app that shows how much time RedLnx has given you back compared with editing by hand, alongside four zeros that name what did not leave your machine: files uploaded, faces indexed, coordinates shared, servers contacted. The number lives in user_profile.json on your disk and is never transmitted.
What is the Kickstarter for?
A one-time pre-launch campaign to cover code-signing, hosting, hardware test costs, and development hours. Not a subscription. Not a paywall. Not a tier system. Funded supporters who opt in are listed alphabetically on the Hall of Fame; everyone else still gets the same software, free, forever.
Is federated training planned?
Long term, yes. The vision is federated training on free and open datasets, never on user-uploaded RAWs, with differential-privacy weight aggregation. Clustered local inference across a photographer's own LAN is also on the long-term map. Both are Phase 4 territory.
[ DOWNLOAD / RELEASES ]
Releases ship the desktop app and the local runtime flow. On Windows, auto mode prefers CUDA when the NVIDIA runtime looks compatible, otherwise it falls back to DirectML, then CPU. On macOS, Metal. On Linux, WebGPU / Vulkan.
CPU is supported too. Acceleration helps some parts of the pipeline more than others, and training quality still depends on the consistency of the source edits.