开源多品牌 RAW 照片编辑器，集调色、AI 增强、智能蒙版与照片合成于一体。

An open-source multi-brand RAW photo editor with color grading, AI enhancement, smart masks, and photo merging.

中文使用说明 · Downloads · Changelog · MIT license

LUMEN RAW is a local desktop photography workspace for Windows and Apple silicon Macs, designed around landscape and travel editing. The application currently has a Chinese interface. No account or cloud service is required to edit photographs; after the runtime assets are installed, editing and bundled AI inference work offline.

Screenshot uses public CC0 test photographs; image credits.

• Natural-language editing (1.5.0): type or say one sentence ("bluer sky, a little warmer, brighten the person") and a language model of your choice returns the slider values, applied as one undoable step with adjustable strength. Use a local runner (Ollama, LM Studio, llama.cpp, any OpenAI-compatible server) or your own cloud API key (OpenAI, Claude, Gemini, DeepSeek, Qwen, Kimi, GLM, SiliconFlow, Volcengine Ark, OpenRouter). Voice input is recognized offline by the bundled SenseVoice-Small model (Chinese and English). A thinking-depth setting (default / off / low / medium / high / max) is translated into each provider's own parameter (DeepSeek thinking + reasoning_effort, OpenAI and Gemini reasoning_effort, Claude output_config.effort, Qwen thinking_budget, Ollama think, ...); with a local qwen3-14b, turning thinking off cut one edit from 11 s to 1.6 s. Cloud use is cheap: with DeepSeek (deepseek-flash), one connection test plus one edit took 2 requests and 2,486 tokens and cost under ¥0.01 (about 2,400 tokens per edit; attaching the preview image, follow-up turns or other models change this).
• Multi-brand RAW: Sony ARW, Canon CR2 / CR3, Nikon NEF, Fujifilm RAF, Panasonic RW2, DNG, and other formats supported by the bundled LibRaw version.
• Non-destructive development: exposure, highlights, shadows, whites, blacks, white balance, per-color HSL, grading, RGB channel curves, snapshots, and portable edit recipes.
• Detail tools: dehaze, clarity, texture, sharpening, noise reduction, spot healing, and clone stamping.
• Masks: brush and gradient masks, luminance ranges, contiguous color selection, and local models for sky, people, subject, background, and relative-depth foreground selections.
• AI enhancement: Real-ESRGAN super-resolution and DRUNet / NAFNet / FFDNet denoising in dedicated preview dialogs, creating editable DNG copies in the filmstrip.
• Photo merging: focus stacking, Mertens exposure fusion with three deghosting strengths, and spherical panoramas capped at 200 megapixels.
• Photo workflow: original-resolution zoom, confirmed crop view, multi-select filmstrip, edit synchronization, and metadata watermark borders.
• Export: JPEG, PNG, 16-bit TIFF, and 16-bit linear RGB DNG. General processing/output limits are 400 megapixels, subject to available memory and disk space.

Get the Windows x64 installer or portable ZIP from Releases. Windows 10 22H2 or Windows 11 is recommended. Portable builds include Python, ExifTool, fonts, and all nine ONNX models; extract the whole folder and keep _internal next to LumenRAW.exe.

For existing installations, save your album and close the previous version before upgrading. Release binaries are currently unsigned.

LumenRAW-1.5.0-macOS-arm64.dmg runs on M1 and newer Macs with macOS 15 Sequoia or later (the PySide6 6.11 bindings are built for macOS 15; every Apple silicon Mac can update to it). Features, including natural-language editing and offline voice input, and the project / album formats match the Windows version.

• Install: open the DMG and drag LUMEN RAW to Applications. The app is ad-hoc signed, not notarized by Apple: the first launch is blocked; open System Settings → Privacy & Security and click Open Anyway (or run xattr -dr com.apple.quarantine "/Applications/LUMEN RAW.app").
• GPU: the pointwise development stages that run on DirectML on Windows (white balance, exposure, tone zones, HSL, curves, grading, ...) run as Metal compute kernels; AI super-resolution, denoising and automatic masks run through ONNX Runtime's Core ML execution provider on the Apple GPU instead of Windows ML / TensorRT for RTX. Both are checked against the CPU on first use and fall back to it on failure. On an M1 Pro a 1600 px preview develops in about 3–5 ms (NumPy: ~760 ms) and the restoration models run 8–10x faster than on the CPU.
• Mac conventions: ⌘ shortcuts, Option-click to set the clone source, two-finger pan, pinch to zoom, Finder "Open With" and Dock drops. Logs are in ~/Library/Logs/LUMEN RAW; ExifTool runs with the system Perl.
• Voice and API keys: the first voice instruction asks for microphone access (speech is recognized on the Mac, nothing is uploaded). Cloud API keys are kept in one item of the login keychain ("LUMEN RAW"); after an update macOS asks once whether the new build may read it — choose Always Allow.
• From source: with Python 3.12, ./run-source.command runs the editor and ./build-macos.command runs the tests, the Metal / Core ML check, PyInstaller and the DMG build (output in .publish/v<version>/macos/, e.g. v150).

• NVIDIA RTX: on Windows 11 24H2+ with a GeForce RTX 30-series or newer GPU, AI super-resolution, denoising and automatic masks use NVIDIA TensorRT for RTX, downloaded once through the Windows ML execution-provider catalog. Other PCs keep DirectML, then CPU.
• Crash-safe AI: neural models run in a separate worker process. If a GPU driver or execution provider crashes, the editor keeps running, retries the tile on the next device, and remembers the GPU + driver combination.
• ONNX Runtime is now the Windows ML build (onnxruntime-windowsml 1.30, CPU and DirectML built in).

• Faster previews: the develop pipeline caches each stage (repair/base, tone, spatial detail, color), so a slider only re-renders the stages after it. Local adjustments are computed only inside each mask's padded bounding box. Both are pixel-identical to a full re-render.
• More work on the GPU: on Windows, DirectML now runs saturation/vibrance, 8-color HSL, RGB and channel curves, monochrome and color grading in addition to tone. Without spatial detail tools, tone and color run as one GPU pass. The graphs are generated at runtime without the onnx package.
• Explicit work state: loading, export, AI/merge, mask inference, previews, detail rendering and thumbnails share one conflict table and one job scheduler.
• Diagnostic logs are written to %LOCALAPPDATA%\LUMEN RAW\logs. LUMEN_COMPUTE=cpu forces the CPU; CuPy is experimental and opt-in (LUMEN_EXPERIMENTAL_CUPY=1).

The installer keeps the original AppId, so it upgrades an existing 1.2.x installation in place. Project and album formats are unchanged.

On Windows, LUMEN RAW uses AMD DirectML for pointwise development and the ONNX models. It picks the DXGI adapter with the most dedicated VRAM, checks that model nodes actually execute on GPU, and falls back to CPU when needed. The center of the bottom bar shows GPU⚡ or CPU⚡（thread count）. RAW decoding and spatial tools (clarity, texture, dehaze, sharpening, noise reduction, retouching) still use CPU.

For a Python 3.12 x64 source environment, install the normal dependencies and runtime assets as below, then switch the ONNX Runtime distribution:

.\.venv\Scripts\python.exe -m pip uninstall -y onnxruntime onnxruntime-directml
.\.venv\Scripts\python.exe -m pip install -r requirements-directml.txt
.\.venv\Scripts\python.exe main.py

The installer and portable ZIP are available from Releases and are rebuilt locally in .publish/v131/packages/, with SHA-256 values in SHA256SUMS.txt. They are unsigned. build-directml.ps1 verifies actual GPU node execution before freezing the application; installer.iss is the Inno Setup recipe. NVIDIA CUDA still uses the separate requirements-gpu.txt flow and has not been validated here. See TEST_REPORT.md for packaged ARW workflow checks.

Use Python 3.12 x64 on Windows. Git contains application code, tests, model metadata, and license notices. Large runtime resources are versioned as a Release asset to keep the repository small.

git clone https://github.com/hety0211/lumen-raw.git
cd lumen-raw
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe tools/fetch_assets.py
.\.venv\Scripts\python.exe main.py

fetch_assets.py downloads the versioned runtime ZIP and verifies the archive and every extracted file against SHA-256 values in runtime-assets.json. It restores the ONNX models, Noto font, ExifTool, and related notices. The initial download is approximately 600 MB; no model download is needed while editing afterward.

To install resources from an already downloaded Release asset:

.\.venv\Scripts\python.exe tools/fetch_assets.py --archive C:\Downloads\LumenARW-1.2.1-RuntimeAssets.zip
.\.venv\Scripts\python.exe tools/fetch_assets.py --verify-only

After the initial resource setup, run-source.cmd can create/install the Python environment and start the application. See MODEL.md for model origins, hashes, conversions, and limitations.

Select 2–32 photographs with Ctrl / Shift in the filmstrip, then right-click → 合成 (Merge). Each method opens its own preview/progress dialog. The default reference is the image with the most original pixels; equal sizes select the middle frame. You can override the reference manually.

Results are named reference-景深合成.dng, reference-HDR堆栈.dng, or reference-全景合成.dng, with a numeric suffix on collision, and automatically added to the filmstrip. Deleting a filmstrip entry removes it from the album, not from disk.

HDR uses OpenCV Mertens exposure fusion, not radiance-map HDR. Focus stacking selects sharp regions at several scales. Panoramas use SIFT, camera optimization, spherical projection, exposure compensation, and tiled feather blending. Motion, parallax, focus breathing, and transparent edges can still produce artifacts.

The Windows build uses DirectML on supported AMD GPUs for pointwise development (tone and color) and ONNX inference, with CPU fallback and up to 32 compute threads. The source supports optional ONNX Runtime CUDA and CuPy; CUDA has not been validated on NVIDIA hardware in this release. See the Chinese setup guide and requirements-gpu.txt before installing GPU dependencies.

AI enhancement operates on developed RGB, not sensor mosaic data, and is not Adobe's RAW enhancement algorithm. Exported DNGs contain developed linear RGB pixels, not lossless copies of the original sensor mosaic. Keep your camera originals and edit recipes.

Version 1.3.0 passed 193 regression tests on the DirectML build (1.2.2: 176; 1.2.1: 166). A controlled large-array panorama test processed 199,197,856 pixels. Photo merge workflows were tested with controlled views derived from one photograph; independently captured real-world brackets and panoramas still need validation. See TEST_REPORT.md for scope and limitations.

.\.venv\Scripts\python.exe -m pip install -r requirements-dev.txt
.\.venv\Scripts\python.exe tools/fetch_assets.py
.\.venv\Scripts\python.exe -m pytest tests -q
powershell -ExecutionPolicy Bypass -File build.ps1

requirements-lock.txt records the original Windows release environment. build.ps1 builds a CPU portable app; build-directml.ps1 builds the tested DirectML app. To compile an installer with Inno Setup 7:

ISCC.exe /DAppBuild=dist\LumenRAW installer.iss

See CONTRIBUTING.md for bug reports and contributions. Please provide only photographs you have permission to share and remove sensitive location/person information before posting samples publicly.

Application code is released under the MIT License. Third-party dependencies, models, fonts, and ExifTool retain their own licenses; see THIRD_PARTY.md, MODEL.md, and notices in assets/. The application does not bundle official camera/lens brand logos.

Built with Python, PySide6 / Qt, NumPy, OpenCV, rawpy / LibRaw, Pillow, tifffile, and ONNX Runtime, with models from Real-ESRGAN, KAIR / DPIR, NAFNet, MiDaS, U²-Net, SkySeg, and TorchVision.