Releases: hety0211/lumen-raw
Release list
LUMEN RAW 1.5.0
LUMEN RAW 1.5.0
自然语言一键修图: 说一句话或打一句话,例如“去掉雾霾,让天空变蓝、河水更碧绿”,你选择的大模型会按软件的固定参数表返回数值,LUMEN RAW 直接应用,一步即可撤销。模型可以用本机的 Ollama / LM Studio / llama.cpp,也可以填写你自己购买的云端 API。内置离线语音识别,支持中文和英文。仅提供 Windows 版本(macOS 用户请继续使用 1.3.1 macOS 版)。工程(.lumen)与选片集(.lumenalbum)格式不变,安装版可直接覆盖安装 1.2.x – 1.4.x。
Natural-language editing: say or type one sentence ("clear the haze, make the sky blue and the river greener") and the language model of your choice returns values for LUMEN RAW's fixed parameter table, applied as one undoable step. Use a local runner (Ollama, LM Studio, llama.cpp) or your own cloud API key. Offline voice input for Chinese and English is built in. Windows only (macOS users stay on the 1.3.1 macOS release). Project and album formats are unchanged; the installer upgrades 1.2.x – 1.4.x in place.
下载 / Downloads
- LumenRAW-1.5.0-Setup.exe:Windows x64 安装版,当前用户安装,无需管理员权限 / Windows x64 installer (per-user).
- LumenRAW-1.5.0-Windows.zip:完整离线便携版,解压后运行
LumenRAW-Windows\LumenRAW.exe/ Offline portable build. - SHA256SUMS.txt:下载校验值 / Download checksums.
安装包比 1.4.1 大约 160 MB,主要是随包的离线语音模型(239 MB,压缩后)。/ The installer is about 160 MB larger than 1.4.1, mostly the bundled offline speech model.
更新内容 / Changes
- 自然语言输入: 左侧面板“预设”和“快照”之间新增“自然语言输入”。输入指令后按 Enter 或“一键修图”;结果区列出模型的说明和每项参数变化,“应用强度”可在 0–150% 之间调节,“撤销本次”恢复修图前,“重新生成”从同一起点再请求一次。同一张照片保留最近三轮对话,可以接着说“再暖一点”。
- AI 能调的参数: 曝光、对比度、亮部、暗部、白色、黑色、冷暖、色调、饱和度、自然饱和度、去薄雾、清晰度、纹理、锐化、降噪;八色 HSL、三段色彩分级、暗角与颗粒、黑白、RGB 曲线;还能自动建立天空 / 主体 / 人物 / 背景 / 近景 AI 蒙版和渐变蒙版做局部调整。模型返回的每个数值都会校验并限制在有效范围内,不会产生无效工程。
- 本地或云端模型: 本地支持 Ollama、LM Studio、llama.cpp,以及 vLLM / Jan / LocalAI 等 OpenAI 兼容服务,无需 API Key,数据不出本机;云端支持 OpenAI、Anthropic Claude、Google Gemini、DeepSeek、阿里云百炼(通义千问)、Kimi、智谱 GLM、硅基流动、火山方舟(豆包)、OpenRouter 和自定义接口。默认地址、模型与参数按各家当前文档设置。API Key 只保存在本机,Windows 上使用 DPAPI 加密。使用云端时只发送指令、当前参数、照片统计和拍摄参数(不含 GPS),可选附带一张长边 768 像素的预览图。
- 思考强度: 设置中可选默认 / 关闭 / 低 / 中 / 高 / 最高,自动换算成各家的参数(DeepSeek
thinking/reasoning_effort、OpenAI 与 Geminireasoning_effort、Claudeoutput_config.effort、通义千问与硅基流动thinking_budget、Ollamathink等);服务不接受时自动去掉重试。修图参数通常“低”或“关闭”就够:本机 LM Studio + qwen3-14b 一次修图,默认 11.3 秒,关闭思考 1.6 秒。 - 用量很小: 每次修图后状态栏显示 token 用量。实测 DeepSeek 测试连接一次加修图一次,共 2 次请求、2,486 个 token,花费不到 ¥0.01。
- 离线语音输入: 点“语音”或按
Ctrl+Shift+Space说话,停顿约 1 秒自动结束,识别后自动修图。语音识别使用随包的开源 SenseVoice-Small(FunAudioLLM),中文、英文和中英混说都能识别,完全离线,6 秒语音约 0.1 秒识别完成。
English summary: a new “自然语言输入” (natural-language input) tab sends your instruction, the photo's current parameters and its statistics to a local or cloud language model and applies the returned values as one step, with adjustable strength, undo and regenerate. Local runners and eleven cloud services (including a custom endpoint) are supported with their current base URLs, models and thinking controls; a thinking-depth setting maps to each provider's own parameter. Token usage is shown after each edit. Voice instructions are recognized offline by the bundled SenseVoice-Small model.
验证 / Validation
- Windows x64,AMD Radeon RX 9070 XT(DirectML):完整回归 290 项通过;便携版与安装包解包版冒烟测试通过,包括随包语音模型、麦克风枚举、TLS 和一次经本地 OpenAI 兼容服务的修图。
- 真实模型:本机 LM Studio 运行 qwen3-14b 与 gemma-4-e4b 完成修图(含天空 AI 蒙版),qwen3-14b 一次 10–20 秒;Windows 语音合成的中文与英文指令逐字识别正确,并经语音自动修图。
- 云端服务按各家文档接入,作者实测了 DeepSeek;其他云端服务欢迎反馈
%LOCALAPPDATA%\LUMEN RAW\logs中的日志。
说明 / Notes
- 语音模型 SenseVoice-Small 遵循 FunASR 模型开源协议 1.1(见
assets/SenseVoice-LICENSE.txt)。/ The speech model is distributed under the FunASR Model License 1.1. - 升级前请保存选片集并关闭旧版。安装包尚未代码签名。/ Save your album and close the previous version before upgrading. Binaries are unsigned.
SHA-256:
4dda735c35a7fcb0fdd000bc4233801626a2b7f4027a5c13e65924fb040db936 LumenRAW-1.5.0-Setup.exe
00b83986542265c2b25b9a68b5d082c442652d25ca5ead4429a605722ceb8a6d LumenRAW-1.5.0-Windows.zip
LUMEN RAW 1.4.1
LUMEN RAW 1.4.1
本版本包含 1.4.0 与 1.4.1 两次更新:高像素照片放大查看不再“未响应”,曝光曲线改为无极调节,新增批量导出,并支持尼康“高效率”NEF 与佳能 HDR PQ CR3。仅提供 Windows 版本(macOS 用户请继续使用 1.3.1 macOS 版)。工程(.lumen)与选片集(.lumenalbum)格式兼容,安装版沿用原 AppId,可直接覆盖安装 1.2.x / 1.3.x。
This release contains the 1.4.0 and 1.4.1 updates: zooming into high-resolution photos no longer freezes the window, the exposure curve becomes stepless, batch export is added, and Nikon "High Efficiency" NEF and Canon HDR PQ CR3 files are supported. Windows only (macOS users stay on the 1.3.1 macOS release). Project and album formats remain compatible; the installer upgrades 1.2.x / 1.3.x in place.
下载 / Downloads
- LumenRAW-1.4.1-Setup.exe:Windows x64 安装版,当前用户安装,无需管理员权限 / Windows x64 installer (per-user).
- LumenRAW-1.4.1-Windows.zip:完整离线便携版,解压后运行
LumenRAW-Windows\LumenRAW.exe/ Offline portable build. - SHA256SUMS.txt:下载校验值 / Download checksums.
1.4.1 更新内容 / Changes in 1.4.1
- 曝光曲线无极调节: 曲线上任意位置都可以上下拖动,曲线只在该亮度附近平滑弯曲,指针下的点准确跟随,远处亮度不变;不再只有四个固定点。黑色/暗部/亮部/白色滑块中,只有拖动位置所在范围的一两个朝拖动方向联动(以前提亮暗部 0.08 就会把黑色、暗部推到 99、84,压低亮部会让纯白大幅下降)。滑块无法表达的局部形状记录为一条精细曲线,随工程、撤销、快照和预设保存;双击曲线全部还原。旧工程的显示与渲染不变。
- 尼康“高效率 / 高效率★” NEF 可以打开: 内置 LibRaw 不支持这种压缩,现在改用文件内嵌的全尺寸 JPEG(Z8 为 8256 × 5504)打开、编辑和导出,界面标注“内嵌 JPEG”(8 位,已应用相机白平衡与风格)。需要原始传感器数据请在相机中使用无损压缩 RAW。
- 佳能 HDR PQ 的 CR3: EOS R5 Mark II 等以 HDR PQ 拍摄的 CR3 内嵌 HEVC 预览,此前图集没有缩略图、显影基准退回“标准亮度”。现在解码该预览并由 PQ 转为 SDR,缩略图和“相机参考显影”恢复正常。
- 批量导出: 底部图集多选后右键“批量导出 N 张…”。可选 JPEG / PNG / TIFF / DNG、质量、原始尺寸或长边 4000 / 2560 / 1600 像素;每张使用各自的调色、裁切、蒙版、修复与水印,后台逐张生成、可取消,同名文件自动编号、不覆盖。实测四张 4500 万–6100 万像素 RAW 全尺寸导出共 29.5 秒,期间界面保持响应。
1.4.0 更新内容 / Changes in 1.4.0
- 高像素照片放大不再“未响应”: 以前放大超过预览后,每次调整都会在界面线程上处理整张原图(6100 万像素每份约 700 MB)。Sony A7R V 实测每次调整、前后对比、按住看原片、开关溢出提示都会让界面卡住 5–6.5 秒,Windows 因此显示“未响应”。现在界面线程不再处理原图分辨率数据,同样操作界面最长停顿约 40 ms,每次调整到细节就绪约 0.4–1 秒(以前约 7–18 秒),峰值内存 5.3 GB → 3.0 GB。
- 只渲染看得见的部分: 放大后只计算屏幕上可见的区域,并按缩放程度选用原图、1/2、1/4… 分辨率;结果切成 512 像素的块缓存,平移只补算新露出的块。与整幅渲染逐像素一致。
- 拖动时不算原图细节: 按住滑块、曲线或拖动画面期间只更新预览,松开后再计算细节;参数或画面一变,正在进行的细节计算立即放弃,不再挡住预览。
- 内存: RAW 全尺寸解码少一份约 700 MB 的临时副本;导出窗口的“中央细节预览”只计算中央区域。
English summary: in 1.4.0 the editor renders only the visible part of a zoomed photo, in 512-pixel tiles at the matching resolution level, entirely off the GUI thread — on a 61 MP Sony A7R V the longest GUI stall per edit drops from 5–6.5 s to about 40 ms and peak memory from 5.3 GB to 3.0 GB. In 1.4.1 the exposure curve can be dragged anywhere and bends locally while the black/shadow/highlight/white sliders of that range stay linked; Nikon High Efficiency NEF files open from their full-size embedded JPEG; Canon HDR PQ CR3 files get thumbnails and a camera-referenced develop curve from their HEVC preview; and selected filmstrip photos can be batch-exported with their own edits.
验证 / Validation
- Windows x64,AMD Radeon RX 9070 XT(DirectML):完整回归 259 项通过;三张 GPU 调色图确认在 GPU 执行;便携版与安装包解包版冒烟测试均为 DirectML;安装包解包后 2736 个文件与便携版逐个一致。
- 实拍样片(raw.pixls.us,CC0):Sony A7R V(6100 万像素)、Nikon Z8(无损与高效率★)、Canon EOS R5 Mark II(HDR PQ),另以 1.5 亿像素 TIFF 做压力测试;安装包解包后的程序在这些文件上完成 100% 细节、AI 天空蒙版、裁切、同步、DNG 导出与选片集流程,原片不变。
- 未在 16 GB 内存或集成显卡的电脑上实测,欢迎反馈
%LOCALAPPDATA%\LUMEN RAW\logs中的日志。
说明 / Notes
- 升级前请保存选片集并关闭旧版。安装包尚未代码签名。/ Save your album and close the previous version before upgrading. Binaries are unsigned.
SHA-256:
4871aa32e5464aa669cbeb7c09f5136fd5922baeab32401e204e25a322f033b5 LumenRAW-1.4.1-Setup.exe
ab91519d50213bacc9407470fa0e80782c722dd1de73136274226bf478b0086d LumenRAW-1.4.1-Windows.zip
LUMEN RAW 1.3.1 · macOS(Apple 芯片)
LUMEN RAW 1.3.1 · macOS(Apple 芯片)
LUMEN RAW 的首个 macOS 版本,基于 1.3.1。编辑功能以及工程(.lumen)、选片集(.lumenalbum)格式与 Windows 版相同。
The first macOS build of LUMEN RAW, based on 1.3.1. Editing features and the project / album formats match the Windows version.
下载 / Downloads
- LumenRAW-1.3.1-macOS-arm64.dmg:Apple 芯片(M1 及更新)Mac,需要 macOS 15 Sequoia 或更新版本(所有 Apple 芯片 Mac 均可免费升级)/ Apple silicon Macs (M1 or newer) with macOS 15 Sequoia or later.
- SHA256SUMS-macOS.txt:下载校验值 / Download checksum.
Windows 安装版与便携版见 LUMEN RAW 1.3.1。/ Windows packages are in the 1.3.1 release.
安装 / Install
- 双击 DMG,把 LUMEN RAW 拖到「应用程序」文件夹。
- 安装包为临时(ad-hoc)签名,没有 Apple 开发者证书公证,首次打开会被系统拦截:点「完成」,打开「系统设置 → 隐私与安全性」,在页面下方点「仍要打开」并确认,之后即可正常使用。也可以在终端执行
xattr -dr com.apple.quarantine "/Applications/LUMEN RAW.app"。
Drag LUMEN RAW to Applications. The app is ad-hoc signed and not notarized by Apple, so the first launch is blocked: open System Settings → Privacy & Security and click Open Anyway (or run the xattr command above).
更新内容 / Changes
- Metal 逐像素显影: Windows 上由 DirectML 运行的白平衡、曝光、光影分区、对比度、饱和度、八色 HSL、曲线、黑白与色彩分级,在 Mac 上由 Metal 计算着色器在 GPU 上运行;首次使用时与 CPU 参考实现自检,失败自动回退 CPU。
- Core ML AI: AI 超分、去杂色和自动蒙版通过 ONNX Runtime 的 Core ML 在 Apple GPU(Metal)上运行,代替 Windows ML / TensorRT for RTX / DirectML;修复类模型按分块尺寸编译,比 CPU 快 8–10 倍。AI 在独立进程中运行,崩溃时自动回退 CPU。
- Mac 操作习惯: ⌘ 快捷键,Option + 单击设置仿制取样点;触控板双指平移、捏合缩放、双指轻点两下适应窗口,鼠标滚轮缩放;可从 Finder「打开方式」或拖到程序坞图标打开照片、工程与选片集。
- 其他: ExifTool 13.59 由系统自带 Perl 运行;日志位于
~/Library/Logs/LUMEN RAW;Core ML 产生的临时模型与从不复用的运行时缓存会自动清理,不会持续占用磁盘。
English: pixel development runs as Metal compute kernels and the AI models run through Core ML on the Apple GPU, replacing Windows ML, TensorRT for RTX and DirectML. The Metal kernels are checked against the CPU reference on first use; both fall back to the CPU on failure, and AI runs in a separate process. The app follows Mac conventions (⌘ shortcuts, Option-click, trackpad pan and pinch zoom, Finder "Open With"), runs ExifTool with the system Perl, and cleans up Core ML's temporary models and caches.
验证 / Validation(Apple M1 Pro,macOS 27)
- 回归测试 219 项通过;同一份源码在 Windows 上 211 项通过,Windows 版行为不变。/ 219 regression tests pass on macOS, 211 on Windows.
- 用 Sony A7 III 原片在源码版、打包版(无窗口与原生窗口)以及直接从挂载的 DMG 中完成编辑、AI 超分预览与导出;整图 AI 去杂色(DRUNet)39 秒,2× 超分(Real-ESRGAN x4plus,输出 5296 × 7872)451 秒。/ End-to-end checks with a Sony ARW from source, the app and the mounted DMG; full-image DRUNet denoise 39 s, 2× Real-ESRGAN 451 s.
| 1600 像素预览 / preview | Metal | CPU |
|---|---|---|
| 含清晰度/纹理/锐化,两个蒙版 / with detail tools, two masks | 88 ms | 656 ms |
| 无空间细节工具 / pointwise only | 31 ms | 592 ms |
| AI 模型 / model(单个分块 / per tile) | Core ML | CPU |
|---|---|---|
| Real-ESRGAN x4plus 192² | 355 ms | 3476 ms |
| DRUNet 320² | 103 ms | 957 ms |
| NAFNet 320² | 49 ms | 416 ms |
说明 / Notes
- 仅支持 Apple 芯片,不支持 Intel Mac;目前只在 macOS 27 上实测。/ Apple silicon only; tested on macOS 27.
- 首次 AI 运算需为每种分块尺寸编译 Core ML 模型,约 0.1–2 秒。/ The first AI run compiles Core ML models per tile size (0.1–2 s).
SHA-256:
465bddb84f3bac1a5a47661ca6df68d734566fa2283ce7837eb971366d6cdd79 LumenRAW-1.3.1-macOS-arm64.dmg
LUMEN RAW 1.3.1
LUMEN RAW 1.3.1
面向 NVIDIA 显卡的稳定性与加速更新。工程(.lumen)与选片集(.lumenalbum)格式不变;安装版沿用原 AppId,可直接覆盖安装 1.2.x / 1.3.0。
A stability and acceleration update for NVIDIA GPUs. Project and album formats are unchanged; the installer upgrades 1.2.x / 1.3.0 in place.
下载 / Downloads
- LumenRAW-1.3.1-Setup.exe:Windows x64 安装版,当前用户安装,无需管理员权限 / Windows x64 installer (per-user).
- LumenRAW-1.3.1-Windows.zip:完整离线便携版,解压后运行
LumenRAW-Windows\LumenRAW.exe/ Offline portable build. - SHA256SUMS.txt:下载校验值 / Download checksums.
更新内容 / Changes
- NVIDIA 显卡使用 TensorRT for RTX: Windows 11 24H2 及以上、GeForce RTX 30 系及更新显卡,AI 超分、去杂色和自动蒙版通过 Windows ML 使用 NVIDIA TensorRT for RTX。首次使用时由 Windows 下载该组件,之后离线可用。其他电脑仍使用 DirectML,最后回退 CPU;调色计算仍使用 DirectML。
- AI 不再导致闪退: 神经网络模型在独立后台进程中运行。显卡驱动或执行设备崩溃时编辑器继续运行,自动换下一种设备(TensorRT → DirectML → CPU)重算,并按“显卡 + 驱动版本 + 模型”记住不兼容的组合(
%LOCALAPPDATA%\LUMEN RAW\gpu-compat.json),更新驱动后会重新尝试。这修复了部分 RTX 50 系显卡在 1.3.0 中 AI 去杂色/超分进行到 15% 时闪退的问题。 - 运行库: ONNX Runtime 改为 Windows ML 版本(
onnxruntime-windowsml1.30,内置 CPU 与 DirectML)。 - 诊断: 日志记录 Windows 版本、各显卡与驱动版本及每个模型实际使用的执行设备;AI 进程另有
lumen-worker.log/crash-worker.log。 - 原图缓存修正: 约 2200 万像素以上的照片在 100% 查看时,分阶段缓存现在也会生效(内存允许时)。
- 更正 1.3.0 说明: 1.3.0 发布说明中的“改 HSL 23–40 ms”来自测速脚本的缓存命中;修正后的测速在 RX 9070 XT 上约为 56–93 ms(1600 像素预览,视是否使用细节工具而定)。
English: on Windows 11 24H2+ with a GeForce RTX 30-series or newer GPU, AI models use NVIDIA TensorRT for RTX through the Windows ML catalog. Neural models now run in a separate worker process, so a GPU driver crash no longer closes the editor; the tile is retried on the next device and the GPU/driver combination is remembered. ONNX Runtime is now the Windows ML build (1.30).
验证 / Validation
- AMD Radeon RX 9070 XT(DirectML):完整回归 200 项通过;三张 GPU 调色图确认在 GPU 执行;源码、便携版和安装包解包版均用 Sony ARW 完成编辑、AI 超分预览(独立进程,DirectML)与 TIFF 导出;安装包解包后 2736 个文件与便携版逐个一致。
- NVIDIA GeForce RTX 5080(社区测试): AI 超分、去杂色全部正常完成,实际使用
NvTensorRtRtxExecutionProvider(TensorRT for RTX);默认高画质超分(Real-ESRGAN x4plus)约每秒 50–60 个分块。每个分块覆盖原图 128 × 128 像素,与相邻分块重叠 32 像素(步进 96 像素),推理输入含 32 像素上下文、最大 192 × 192,4× 输出最大 768 × 768;按此估算,一张 2400 万像素(6000 × 4000)照片约 2600 个分块,约 45–55 秒。 - 其他 NVIDIA 显卡如 TensorRT for RTX 不可用,程序会回退到 DirectML 或 CPU 而不闪退;欢迎反馈
%LOCALAPPDATA%\LUMEN RAW\logs中的日志。 - Community test on a GeForce RTX 5080: AI super-resolution and denoising completed using TensorRT for RTX (
NvTensorRtRtxExecutionProvider), about 50–60 tiles per second with the default Real-ESRGAN x4plus model. Each tile covers 128 × 128 source pixels with a 32-pixel overlap (96-pixel stride) and up to 192 × 192 input pixels including context (768 × 768 at 4×) — roughly 2,600 tiles, or 45–55 seconds, for a 24 MP photo. Where TensorRT for RTX is unavailable the app falls back to DirectML or CPU instead of crashing.
说明 / Notes
- 升级前请保存选片集并关闭旧版。安装包尚未代码签名。/ Save your album and close the previous version before upgrading. Binaries are unsigned.
SHA-256:
b8ea02b8f7c173dd9cf9494d663c7e108b34c7b816cc8e7818dd4bc40b4a7139 LumenRAW-1.3.1-Setup.exe
c166f332b10471c75d734c3a68ce8def2b00baa4c3fa3d12bfa8420437e40c32 LumenRAW-1.3.1-Windows.zip
LUMEN RAW 1.3.0
LUMEN RAW 1.3.0
预览更快、更多步骤使用 GPU、后台任务管理更清晰。工程(.lumen)与选片集(.lumenalbum)格式不变;安装版沿用原 AppId,可直接覆盖安装 1.2.x。
Faster previews, more GPU work and clearer background task handling. Project and album formats are unchanged, and the installer upgrades 1.2.x in place.
下载 / Downloads
- LumenRAW-1.3.0-Setup.exe:Windows x64 安装版,当前用户安装,无需管理员权限 / Windows x64 installer (per-user, no admin rights).
- LumenRAW-1.3.0-Windows.zip:完整离线便携版,解压后运行
LumenRAW-Windows\LumenRAW.exe,保留_internal文件夹 / Offline portable build; keep_internalnext to the EXE. - SHA256SUMS.txt:下载校验值 / Download checksums.
从源码运行所需的模型、字体和 ExifTool 仍使用 v1.2.1 的 RuntimeAssets,python tools/fetch_assets.py 会自动下载并校验。
Source checkouts keep using the v1.2.1 runtime-asset bundle via python tools/fetch_assets.py.
更新内容 / Changes
- 预览分阶段缓存: 修复/显影、光影、空间细节(清晰度、纹理、锐化、降噪、去薄雾)、色彩四个阶段分别缓存,拖动滑块只重算其后的阶段;蒙版范围与亮度参考也会缓存。结果与不缓存时逐像素一致,切换照片自动失效。
- 蒙版局部计算: 局部调整只处理蒙版覆盖的外接矩形,并按滤镜半径补足上下文,结果与整幅计算一致;正向去薄雾依赖整幅统计,仍整幅计算。
- 更多步骤用 GPU(DirectML): 除光影外,饱和度/自然饱和度、八色 HSL、RGB 与单通道曲线、黑白和三段色彩分级也在 GPU 上运行;未使用空间细节工具时,光影与色彩合并为一次 GPU 处理。计算图在运行时生成,不再依赖
onnx包。 - 计算设备选择集中:
LUMEN_COMPUTE=cpu|dml|cuda|auto可用于排查问题;未经实机验证的 CuPy 改为需LUMEN_EXPERIMENTAL_CUPY=1才启用。 - 显式任务状态与调度器: 读取、导出、AI/合成、蒙版识别、预览、原图细节和缩略图由同一张冲突表和同一个任务队列管理。
- 诊断日志: GPU 回退、任务失败写入
%LOCALAPPDATA%\LUMEN RAW\logs\lumen.log,原生崩溃写入同目录crash.log。 - 安装升级: 覆盖安装时先移除旧版
_internal运行库再写入新文件;版本号统一由lumen/__init__.py维护。
English summary: the develop pipeline now caches each stage so a slider only re-renders what follows it; local adjustments are computed only inside each mask's padded bounding box (pixel-identical to a full render); DirectML also runs saturation/vibrance, 8-color HSL, curves, monochrome and grading, fused with tone when no spatial tool is active; work state and background jobs are handled by one scheduler; diagnostic logs go to %LOCALAPPDATA%\LUMEN RAW\logs.
验证 / Validation
在 Windows x64、AMD Radeon RX 9070 XT(DirectML)上:
- 完整回归 193 项通过(1.2.2 为 176 项)。/ 193 regression tests pass.
- 三张 GPU 计算图均确认在 GPU 上执行节点。
- 源码、冻结 EXE 与安装包解包后的 EXE 用 Sony A7 III ARW 完成完整编辑、AI 细节预览与 TIFF 导出,均使用 DirectML;原片只读。
- 安装包解包后 2727 个文件与便携版逐个 SHA-256 一致。
| 1600 px 预览、两个蒙版 | DirectML | CPU |
|---|---|---|
| 含清晰度等细节工具,完整渲染 | 156 ms | 678 ms |
| 无细节工具,完整渲染 | 75 ms | 594 ms |
| 之后调整 HSL 或蒙版(缓存) | 23–40 ms | 34–60 ms |
说明 / Notes
- 升级前请保存选片集并关闭旧版。/ Save your album and close the previous version before upgrading.
- 安装包尚未代码签名,Windows SmartScreen 可能提示。/ Binaries are unsigned.
- NVIDIA CUDA 尚未实机验证;DNG 输出为已显影的线性 RGB,并非原始传感器数据。/ CUDA is not hardware-validated; exported DNGs contain developed linear RGB.
SHA-256:
f2cb6720a0fa1c9e4426edb9c166873c922f92bb3ca588c3e332d691b7034fc8 LumenRAW-1.3.0-Setup.exe
d74bd248b2c356dd5aaa8de04dc483e9d1ec5bd0a7a481a0ab71835ddca7e872 LumenRAW-1.3.0-Windows.zip
LUMEN RAW 1.2.2
LUMEN RAW 1.2.2
原 Lumen ARW 现更名为 LUMEN RAW,以体现 Sony、Canon、Nikon、Fujifilm、Panasonic 等多品牌 RAW 支持。安装程序沿用原 AppId,新版可读取旧版 .lumen 工程与 .lumenalbum 选片集。
在 Windows 上,基础光影显影、AI 超分、去杂色和智能蒙版可通过 DirectML 使用 AMD GPU;首次推理会确认节点实际在 GPU 上运行,失败时回退 CPU。底部状态栏显示 GPU⚡ 或 CPU⚡(线程数)。RAW 解码与部分空间编辑仍使用 CPU。此次在 Radeon RX 9070 XT 上实测,改名后 176 项自动化检查通过。
下载 LumenRAW-1.2.2-Setup.exe 安装,或解压 LumenRAW-1.2.2-Windows.zip 后运行 LumenRAW-Windows/LumenRAW.exe。使用 SHA256SUMS.txt 核对文件。两种运行包均含离线模型和 ExifTool,安装包尚未代码签名。
English: LUMEN RAW replaces the former Lumen ARW name. This release supports multi-brand RAW files and DirectML GPU processing on supported AMD hardware, with a CPU fallback. Existing project and album files remain readable. The installer and portable ZIP include offline runtime assets. The binaries are unsigned.
Lumen ARW 1.2.1
Lumen ARW 1.2.1
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.
下载 / Downloads
- LumenARW-1.2.1-Setup.exe:Windows x64 安装版 / Windows x64 installer.
- LumenARW-1.2.1-Windows.zip:完整离线便携版,解压后运行 / Offline portable build; extract the entire folder.
- LumenARW-1.2.1-RuntimeAssets.zip:从 GitHub 源码运行所需的九个 ONNX 模型、字体、ExifTool 和许可文本 / Runtime resources for source checkouts.
- SHA256SUMS.txt:下载文件的 SHA-256 校验值 / Download checksums.
GitHub 自动生成的 Source code 压缩包不含大型资源。源码用户运行 python tools/fetch_assets.py 下载并校验资源,或使用 --archive 指定已下载的资源 ZIP。运行包已包含全部资源。应用界面目前为中文。
GitHub's generated Source code archives do not contain large runtime resources. Run python tools/fetch_assets.py after cloning, or use --archive with the downloaded resource ZIP. Installer and portable builds already include all runtime resources. The interface is currently Chinese.
1.2.1 更新 / Changes
- 图集右键删除和合成;删除只移除条目,不删除磁盘原片。 / Filmstrip removal preserves original files.
- 独立窗口提供景深合成、Mertens HDR 堆栈及低/中/高去伪影、球面全景合成。 / Dedicated dialogs for focus stacking, Mertens exposure fusion with three deghosting strengths, and spherical panoramas.
- 默认以原始像素最多的照片为参考,合成 DNG 自动命名并加入图集。 / The largest original image is the default reference; generated DNG copies are named and added automatically.
- 全景画布上限 2 亿像素;保留此前 4 亿像素一般处理上限及最多 32 个计算线程。 / Panoramas are capped at 200 MP; the general 400 MP limit and up to 32 compute threads remain.
验证与限制 / Validation and limitations
桌面发布通过 166 项回归检查,三种合成均通过打包后 EXE 的完整流程。受控大图测试实际计算 199,197,856 像素全景。GitHub 源码另加入资源下载与校验测试;测试范围见仓库 TEST_REPORT.md。
The desktop release passed 166 regression checks and packaged EXE workflows for all three merge methods. A controlled test computed a 199,197,856-pixel panorama. The GitHub source additionally tests resource installation and verification; see TEST_REPORT.md.
合成使用 CPU,随包 AI 使用 ONNX CPU。CUDA 接口尚未实机验证。实拍多张序列仍待验证;受控测试不能保证所有运动或视差场景无接缝。DNG 输出为已显影的线性 RGB,并非原始传感器马赛克。
Merging and bundled AI use CPU. CUDA interfaces remain unvalidated on NVIDIA hardware. Real captured multi-photo sequences still need testing; motion and parallax can leave artifacts. Exported DNGs contain developed linear RGB, not original sensor mosaics.
软件源码采用 MIT,第三方组件和模型遵循各自许可。安装包尚未代码签名。升级前请保存选片集并关闭旧版。
Application code is MIT-licensed; third-party components and models retain their licenses. Binaries are unsigned. Save your album and close the previous version before upgrading.