1.6.0

New

• Chromaforge too. AI Remove now works in Chromaforge (once
  "LightCraft Classic") just as in LightCraft, in one of the two or in both at once. The installer asks which app to set up.
• Both open at once. LightCraft uses port 7980 and Chromaforge port 7981. Ctrl/Cmd+Shift+X opens the photo of the
  app you pressed it in: AI Remove asks each app which photo it has selected.
• Several photos selected in Chromaforge: the active one opens.
• Chromaforge's control channel token: AI Remove reads it from Chromaforge's settings folder.
• The window names the app: "Save to Chromaforge", "Update in Chromaforge"…
• SAM 3 is shared with either app. A RAW opened on its own is read with LightCraft's or Chromaforge's command line.
• --app lightcraft|chromaforge for the window and --auto, and --apps for the installer. The MCP tools take
  app and apps.
• The name is now "AI Remove for LightCraft and Chromaforge" (installer, Start menu, Settings > Apps). An update
  keeps everything you have.

Fixed

• Chromaforge passes every selected photo to its external editor in one command: AI Remove used to refuse that and not
  open.

Download

System	Package	Then
Windows 10/11	AIRemove-1.6.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.6.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.6.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.6.0-all-platforms.zip

The installer asks which app to set up: LightCraft, Chromaforge or both. Then restart that app once, select a photo
and press Ctrl+Shift+X (Cmd+Shift+X on a Mac). Or open AI Remove from the Start menu, Applications or the app
menu and choose a photo: it works on its own too (a RAW needs LightCraft or Chromaforge installed, a JPEG, PNG, TIFF,
WebP, HEIC or AVIF does not).

Already installed? Run the 1.6.0 installer again on the same folder. It updates the app and keeps the models
you have, without downloading them again.

Checksums: SHA256SUMS has the SHA-256 of each package. To check a download, in its folder:

• Linux: sha256sum -c SHA256SUMS --ignore-missing
• macOS: shasum -a 256 -c SHA256SUMS --ignore-missing
• Windows (PowerShell): Get-FileHash AIRemove-1.6.0-windows.zip, then compare with SHA256SUMS

The packages were built and checked by GitHub Actions from commit 96b29bd (the run).

Licences

The models have their own licences:

• LaMa: Apache-2.0
• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• RORem: Apache-2.0, on SDXL Inpainting (CreativeML Open RAIL++-M)
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• Qwen-Image 2.1 Turbo: Qwen Research License, non-commercial; its text encoder Qwen3-VL: Apache-2.0
• DAI: Apache-2.0, a fine-tune of Stable Diffusion 2 (CreativeML Open RAIL++-M)

For paid work, use LaMa or RORem, and DAI for reflections. See THIRD-PARTY-NOTICES.md for details.

What's Changed

• Chromaforge support: AI Remove works in it too; the installer asks which app to set up by @AndrianLeah in #38
• 1.6.0 by @AndrianLeah in #39

Full Changelog: v1.5.0...v1.6.0

1.5.0

New

• Remove reflections, like Lightroom's: a photo through glass (a shop window, a display case, a car window) without
  the light the glass reflects (#35, #36).
  • Choose DAI in the menu and press Remove reflections. With nothing selected it works on the whole photo;
    with a selection, only there, fading out around it.
  • Every detail stays: the reflection's light is taken off the RAW's own values, and nothing is redrawn.
  • Strength, under the result, sets how much of the reflection goes, from 0 to 150 %. It starts at 75 %, the
    closest to Lightroom. The photo follows while you drag, and the project and the DNG keep it.
  • About 6 s on a GPU for a whole photo. A photo with a reflection removal reopens in about 3 s.
  • DAI is a new optional model (5.7 GB). It is not for a sharp, bright reflection (trees mirrored in a window, the
    sky in a puddle), which is only partly taken away.
  • Headless: airemove-cli --auto dai (dai100 for 100 %), and the MCP tool remove_reflections.
• Describe: type what to select in English ("people", "car", "sign") and SAM 3 selects every instance in the
  photo (#28).
  • Each instance is its own object: a right-click drops one you want to keep.
  • About 0.35 s each time.
• Split into parts, a model for each (#31, #34).
  • Select everything you want gone in one go. Split into parts numbers the parts on the photo.
  • Choose a model for each part and press Generate, or Generate all for the rest.
  • The parts are sized for the models, so a crowd is no longer one soft fill.
  • Remove selection still does everything at once, with one model.
• Four models in the menu, each saying what it is not for (#32):
  • LaMa: small things on plain backgrounds;
  • RORem, the default: people and objects, and it draws no one back;
  • FLUX Fill: architecture and large areas;
  • Qwen-Image: people and objects on complex backgrounds.
  • Qwen-Image now names everything selected ("the woman and her handbag") and is told what Describe found, so
    nothing selected is left behind.
• Out of the menu (#32):
  • SDXL: RORem runs on its files, so it is installed with RORem.
  • OSOR: it only works behind "Include the shadow".
  • ObjectClear is retired. Updating removes its 7.3 GB of files from AI Remove's own models folder; a folder of
    your own that it was linked from is never touched.
• The window fits any screen (#29). The side panel scrolls, with Save and Close always in sight. The top bar takes
  two rows when it is narrow. The smallest window is 840 × 520.
• Smaller things (#34):
  • With the Brush and Eraser, Alt+wheel (Option on a Mac) and + − set the size, and Ctrl+wheel (Cmd) zooms.
  • Two variants by default, and the number you choose is remembered.
  • The text field is hidden for models that read no text.

Fixed

• A fill could look tinted in the window, red for example, though it was right (#33). A few stray pixels around
  the hole bent the whole channel's preview curve. The DNG was never affected.

Tested

• Windows 11, AMD Radeon RX 7800 XT:
  • An install updated from 1.4.0, with every model kept. The installer downloaded DAI (5.7 GB).
  • In the window, on copies of Canon photos:
    • reflections removed from a statue behind a shop window: the Strength slider, saved, reopened, and the same DNG
      byte for byte;
    • Describe and Split into parts on a crowded street, each part with its own model;
    • window sizes from 840 × 520 to full screen.
• GitHub Actions, on Windows, macOS 14 and Ubuntu: the removal, split, Describe, tiles, pictures and reflections
  tests, through the real app. The reflections test uses a stand-in for DAI on a synthetic RAW with a known
  reflection. The packages below were built and checked there.
• Not tested yet: DAI on a Mac or on a Linux GPU, and DAI on the processor alone. Reports are welcome.

Download

System	Package	Then
Windows 10/11	AIRemove-1.5.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.5.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.5.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.5.0-all-platforms.zip

Then restart LightCraft once, select a photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac). Or open
AI Remove from the Start menu, Applications or the app menu and choose a photo: it works on its own too (a RAW
needs LightCraft installed, a JPEG, PNG, TIFF, WebP, HEIC or AVIF does not).

Already installed? Run the 1.5.0 installer again on the same folder. It updates the app and keeps the models
you have, without downloading them again. It also asks whether you want DAI, for Remove reflections
(5.7 GB).

Checksums: SHA256SUMS has the SHA-256 of each package. To check a download, in its folder:

• Linux: sha256sum -c SHA256SUMS --ignore-missing
• macOS: shasum -a 256 -c SHA256SUMS --ignore-missing
• Windows (PowerShell): Get-FileHash AIRemove-1.5.0-windows.zip, then compare with SHA256SUMS

The packages were built and checked by GitHub Actions from commit cf9aaa0 (the run).

Licences

The models have their own licences:

• LaMa: Apache-2.0
• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• RORem: Apache-2.0, on SDXL Inpainting (CreativeML Open RAIL++-M)
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• Qwen-Image 2.1 Turbo: Qwen Research License, non-commercial; its text encoder Qwen3-VL: Apache-2.0
• DAI: Apache-2.0, a fine-tune of Stable Diffusion 2 (CreativeML Open RAIL++-M)

For paid work, use LaMa or RORem, and DAI for reflections. See THIRD-PARTY-NOTICES.md for details.

1.4.0

New

• Ordinary pictures too: JPEG, PNG, TIFF, WebP, HEIC and AVIF. They open like RAWs: from LightCraft with
  Ctrl+Shift+X, or on their own (#26).
  • The result is a copy in the same format next to the original, <name>-AIRemove.jpg (.png, .heic…).
  • In LightCraft it is stacked on the original with the same development, as the DNG is.
  • Opened on its own, a picture does not need LightCraft at all.
• What a picture keeps:
  • its colour profile, its EXIF and XMP, its alpha channel, and 16 bits for a 16-bit TIFF;
  • the EXIF loses only its small embedded preview, which would still show what was removed.
• How each format is saved:
  • JPEG, with the original's own compression settings: outside the removal, pixels move by a level or two,
    invisibly;
  • PNG, TIFF and lossless WebP, losslessly;
  • HEIC at quality 70: about a third bigger than an iPhone original, untouched pixels moving by half a level;
  • AVIF at quality 90.
• HEIC support uses pillow-heif, which the installer now adds. Its packages are GPLv2 as a whole; AI Remove does
  not ship them (see THIRD-PARTY-NOTICES.md). AVIF is built into Pillow.
• Window: Open photo… replaces Open RAW…. A picture opened on its own has Save as JPEG (PNG, HEIC…).

Not supported yet: a 16-bit colour PNG, CMYK or animated pictures (the message says what to do instead). HEIC
depth maps and HDR (10-bit) data are not kept.

Tested

• Windows 11, AMD Radeon RX 7800 XT:

  • The 1.4 installer updating an existing install: pillow-heif added, every model kept.
  • In LightCraft, removed, saved and stacked:
    • a JPEG, with its development (+0.5 exposure) carried over to the result;
    • an iPhone HEIC, kept upright.
  • On its own, on a JPEG: every EXIF tag kept, the embedded preview gone.

• GitHub Actions, on Windows, macOS 14 and Ubuntu: the RAW tests as before, plus a removal on each of the six
  picture formats (tools/ci_pictures.py) through the real app. The packages below were built and checked there.

• Not tested yet:

  • HEIC and AVIF results opened in other apps (Photos, Finder);
  • pictures on macOS and Linux beyond what CI checks.

  Reports are welcome.

Download

System	Package	Then
Windows 10/11	AIRemove-1.4.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.4.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.4.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.4.0-all-platforms.zip

Then restart LightCraft once, select a photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac). Or open
AI Remove from the Start menu, Applications or the app menu and choose a photo: it works on its own too (a RAW
needs LightCraft installed, a JPEG, PNG, TIFF, WebP, HEIC or AVIF does not).

Already installed? Run the 1.4.0 installer again on the same folder. It updates the app and keeps the models
you have, without downloading them again.

Checksums: SHA256SUMS has the SHA-256 of each package. To check a download, in its folder:

• Linux: sha256sum -c SHA256SUMS --ignore-missing
• macOS: shasum -a 256 -c SHA256SUMS --ignore-missing
• Windows (PowerShell): Get-FileHash AIRemove-1.4.0-windows.zip, then compare with SHA256SUMS

The packages were built and checked by GitHub Actions from commit 68061bf (the run).

Licences

The models have their own licences:

• LaMa: Apache-2.0
• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• RORem: Apache-2.0, on SDXL Inpainting (CreativeML Open RAIL++-M)
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• ObjectClear: S-Lab License 1.0, non-commercial. Its code is downloaded at install time, not shipped with AI
  Remove.
• Qwen-Image 2.1 Turbo: Qwen Research License, non-commercial; its text encoder Qwen3-VL: Apache-2.0

For paid work, use LaMa, SDXL or RORem. See THIRD-PARTY-NOTICES.md for details.

What's Changed

• Ordinary pictures too: JPEG, PNG, TIFF, WebP, HEIC, AVIF by @AndrianLeah in #26
• 1.4.0 by @AndrianLeah in #27

Full Changelog: v1.3.0...v1.4.0

1.3.0

New

• AI Remove works on its own, too. Open it from the Start menu, Applications or the app menu with LightCraft
  closed (or with no photo selected there), and it asks for a RAW. Open RAW… in the window opens another one.
  • LightCraft must be installed, not open: AI Remove reads the RAW through LightCraft's command line, with its
    default development.
  • Save as DNG writes <name>-AIRemove.dng next to the RAW, without importing it anywhere (#24).
• Double-click a project to open its photo. .airemove files open in AI Remove, and have their own icon: a page
  with the raccoon (#24).
• Qwen-Image 2.1 (optional, 16 GB graphics cards). It first recognises what is selected ("the woman with a
  handbag", "the floor lamp"), then removes it by name and redraws the scene behind it. It is for people and objects
  in front of steps, columns or railings.
  • About 30 s per variant on a 16 GB AMD card (#15, #16, #18, #22).
  • Non-commercial licence.
• RORem (optional). It is SDXL Inpainting retrained only to remove. Where SDXL draws a person or an odd object
  back into the hole, RORem draws the background. In 12 test variants it drew nothing back; SDXL drew a person, a
  bag or an owl back in 11 (#21).
  • Commercial use allowed, like SDXL.
  • It adds 4.8 GB to SDXL.
  • SDXL stays the default, because it is better on poles and paths.
• A clearer Generate panel (#19):
  • the model menu shows each model's name, grouped by what it is for, and says when to choose it;
  • ＋ more for #n regenerates a removal with the model chosen above;
  • the text field says what it is for with each model (what to put there, what to remove, or not used);
  • Ctrl+wheel (Cmd+wheel on a Mac) sets the Brush and Eraser size;
  • Full detail says when it applies.
• Reopening a photo with a project is about 5× faster: 37.5 s to 7.7 s with 5 removals, with the same results
  (#20).
• Licences where you choose a model. FLUX's description says it is non-commercial, and the README has a
  Commercial use section (#14).

Fixed

• The installer no longer stops when another program reads its status file at the wrong moment (Windows, #17).
• Qwen-Image: a reply with no name no longer leaves an empty one in the edit and the window (#23).
• RORem ignores typed text everywhere, as it was measured, also with --auto and the MCP server (#23).
• Full detail on a reopened removal shows one progress bar instead of restarting it at 0 % (#23).
• A RAW AI Remove cannot read yet (Fuji X-Trans, an already demosaiced DNG) now says so, instead of an error
  trace (#24).

Tested

• Windows 11, AMD Radeon RX 7800 XT:

  • The 1.3.0 installer updating an existing install, with every model.
  • Opened from LightCraft on a photo with a 5-removal project.
  • Opened on its own and by double-clicking a project, on a CR2.
  • RORem and Qwen-Image there, one after the other, on the same removal.
  • Qwen-Image and RORem measured against SDXL on real photos, written into the RAW and developed by LightCraft.

• GitHub Actions, on Windows, macOS 14 and Ubuntu:

  • the real installer (processor / Metal, LaMa), with a model reused from its own folder;
  • a full removal, and a long removal in squares;
  • the MCP server;
  • the .airemove file type and its icon;
  • the uninstaller.

  The packages below were built and checked there.

• Not tested yet:

  • the bigger models on NVIDIA, Intel Arc or Apple GPUs;
  • opening on its own and the .airemove double-click on macOS and Linux (CI checks that they are registered).

  Reports are welcome.

Download

System	Package	Then
Windows 10/11	AIRemove-1.3.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.3.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.3.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.3.0-all-platforms.zip

Then restart LightCraft once, select a raw photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac). Or open
AI Remove from the Start menu, Applications or the app menu and choose a RAW: it works on its own too, with
LightCraft installed.

Already installed? Run the 1.3.0 installer again on the same folder. It updates the app and keeps the models
you have, without downloading them again.

Checksums: SHA256SUMS has the SHA-256 of each package. To check a download, in its folder:

• Linux: sha256sum -c SHA256SUMS --ignore-missing
• macOS: shasum -a 256 -c SHA256SUMS --ignore-missing
• Windows (PowerShell): Get-FileHash AIRemove-1.3.0-windows.zip, then compare with SHA256SUMS

The packages were built and checked by GitHub Actions from commit f74fe12 (the run).

Licences

The models have their own licences:

• LaMa: Apache-2.0
• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• RORem: Apache-2.0, on SDXL Inpainting (CreativeML Open RAIL++-M)
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• ObjectClear: S-Lab License 1.0, non-commercial. Its code is downloaded at install time, not shipped with AI
  Remove.
• Qwen-Image 2.1 Turbo: Qwen Research License, non-commercial; its text encoder Qwen3-VL: Apache-2.0

For paid work, use LaMa, SDXL or RORem. See THIRD-PARTY-NOTICES.md for details.

What's Changed

• Licences: say what is non-commercial where people choose by @AndrianLeah in #14
• Qwen-Image 2.1 as an optional engine: removes the selection by name (16 GB GPUs) by @AndrianLeah in #15
• Qwen-Image in squares: name once, read every square once per removal by @AndrianLeah in #16
• Installer: a program reading install-status.json no longer stops the install by @AndrianLeah in #17
• Qwen-Image: an oblong area in one pass in its own shape, not in squares by @AndrianLeah in #18
• Window: clearer Generate panel (model names, regenerate a removal, Qwen's field, brush size, Full detail) by @AndrianLeah in #19
• Opening: 37.5 s -> 7.7 s on a 5-removal project, results bit-identical by @AndrianLeah in #20
• RORem as an optional engine: SDXL retrained to remove, draws no one back by @AndrianLeah in #21
• Qwen-Image: 4 steps instead of Turbo's 8, ~30% less per variant by @AndrianLeah in #22
• Cleanup before 1.3.0: the review's findings, one list of models by @AndrianLeah in #23
• Open a RAW on its own, and .airemove projects with a double-click by @AndrianLeah in #24
• 1.3.0 by @AndrianLeah in #25

Full Changelog: v1.2.0...v1.3.0

1.2.0

New

• Long removals stay sharp. Poles, cables and railings used to be filled in one square around the whole object,
  drawn small and enlarged up to 5 times. They are now filled in squares along them, close to the model's own
  resolution. On a pole as tall as the photo, FLUX now gets the capital, the slender column and the tiles behind it
  right, in about a minute per variant (#8).
• Full detail redraws the chosen variant of a long removal at the photo's own resolution, once, in about 2–3
  minutes (#8).
• Thin cords and cables get an automatic LaMa variant, which is the sharpest and fastest fill there, in seconds
  (#8).
• Stop (Esc) while variants are being made. The ones already made are kept (#10).
• A lighter window (#9, #10):
  • the model's description is behind What it does;
  • thumbnails show what was filled;
  • ← / → step through the variants;
  • right-click a variant to delete it.
• The version and a GitHub link next to the title (#12, #13). A note appears in the top bar when a newer version
  is out (#10). AI Remove asks GitHub at most once a day, and sends nothing about you or your photos. You can turn
  the check off.
• An installer that explains each model: what it is for, its download size and its licence. It can also use a
  model you already have in another folder (another AI Remove, a Hugging Face cache, ComfyUI's models) instead of
  downloading it again (#11). Running it again keeps the models you have (#4).

Fixed

• SDXL no longer invents mounds, rocks or animals where the hole meets the sky. It also no longer leaves a ghost
  of the object in plain sky: it now sees the photo upright and keeps the sky as sky (#5). FLUX keeps working on the
  photo as the sensor recorded it, which measured better for FLUX (#6).
• Clipped skies are white, not magenta. Fills in a clipped sky go back into the raw as clipped highlights, with no
  pink patch (#7).
• A selection longer than the photo's short side no longer loses its ends (#8).
• LaMa is much faster after the first fill, because it is loaded only once (#8).

Tested

• Windows 11, AMD Radeon RX 7800 XT: used in LightCraft, including a pole as tall as the photo (squares and Full
  detail).
• GitHub Actions, on Windows, macOS 14 and Ubuntu: the real installer (processor / Metal, LaMa), a model reused
  from its own folder, a full removal, a long removal in squares, the MCP server and the uninstaller. The packages
  below were built and checked there.
• Not tested yet: the bigger models on NVIDIA, Intel Arc or Apple GPUs. Reports are welcome.

Download

System	Package	Then
Windows 10/11	AIRemove-1.2.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.2.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.2.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.2.0-all-platforms.zip

Then restart LightCraft once, select a raw photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac).

Already installed? Run the 1.2.0 installer again on the same folder. It updates the app and keeps the models you
have, without downloading them again.

Checksums: SHA256SUMS has the SHA-256 of each package. To check a download, in its folder:

• Linux: sha256sum -c SHA256SUMS --ignore-missing
• macOS: shasum -a 256 -c SHA256SUMS --ignore-missing
• Windows (PowerShell): Get-FileHash AIRemove-1.2.0-windows.zip, then compare with SHA256SUMS

The packages were built and checked by GitHub Actions from commit 6f18e8b
(the run).

Licences

The models have their own licences:

• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• ObjectClear: S-Lab License 1.0, non-commercial. Its code is downloaded at install time, not shipped with AI
  Remove.

See THIRD-PARTY-NOTICES.md for details.

Full changelog: v1.1.0...v1.2.0

1.1.0

New

• Remove the shadow too. Select a person or an object, press ☀ Include the shadow, and its shadow and
  reflection are added to the selection, in blue. Any model then fills the larger hole. The button appears when
  OSOR or ObjectClear is installed.

• Two optional models that remove an object together with its shadow and reflection, even when you select only
  the object:

  • ObjectClear, about 5 s per variant. It gave the most natural results in our tests. Download: 7.3 GB.
  • OSOR, one step, about 8 s. It runs on FLUX, so it needs FLUX installed. Download: 0.6 GB.

  Both are free for non-commercial use only. The installer asks before downloading either one.

• SDXL is about twice as fast. It now takes 12 steps instead of 25, with the same fills for the same seed: 7.4 s
  per variant instead of 15.4 s on an RX 7800 XT at 1024 px. Every variant after the first saves about 1.3 s more,
  because the photo is prepared only once per removal.

• A clearer model menu. Each entry says what the model is for and how long it takes, for example "Object + shadow ·
  ~5 s each". The model's name and the details are shown under the menu.

Fixed

• The sky's colour no longer leaks into fills where the model drew a wall or grass along the edge.
• Dark, flat areas no longer get a "lace" texture copied from nearby edges.
• From 1.0.1: PyTorch for the right GPU, the Windows uninstaller's exit code and --reinstall.

Download

System	Package	Then
Windows 10/11	AIRemove-1.1.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.1.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.1.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.1.0-all-platforms.zip

Then restart LightCraft once, select a raw photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac).

Already installed? Run the 1.1.0 installer again on the same folder. It updates the app and keeps the models you
have, without downloading them again. To add the new models, answer yes when it offers them. It offers OSOR only if
you answer yes to FLUX, which costs nothing when FLUX is already there.

Tested

• Windows 11 on an AMD Radeon RX 7800 XT: the whole flow on a real raw: selection, Include the shadow, then SDXL,
  OSOR and ObjectClear, with the window inside LightCraft.
• GitHub Actions: the real installer runs on Windows, macOS 14 and Ubuntu (processor / Metal, LaMa only), then the
  app's self-check, the MCP server and the uninstaller.
• Not tested yet: the new models on NVIDIA, Intel Arc or Apple GPUs. Reports are welcome.

Licences

The models have their own licences:

• SAM 3: SAM License
• SDXL: CreativeML Open RAIL++-M
• FLUX.1 [dev] and OSOR: FLUX.1 [dev] Non-Commercial
• ObjectClear: S-Lab License 1.0, non-commercial. Its code is downloaded at install time, not shipped with AI
  Remove.

See THIRD-PARTY-NOTICES.md for details.

1.0.1

A bug-fix release of AI Remove for LightCraft.

Fixed

• PyTorch for your GPU. The installer asked PyPI before PyTorch's own index, so some systems got a PyTorch that
  cannot use their graphics card and the AI ran on the processor: NVIDIA on Windows, AMD on Linux and Intel Arc.
  Linux installs without a GPU also downloaded about 3 GB of NVIDIA libraries they don't need. Macs, NVIDIA on Linux
  and AMD on Windows were not affected.
• Windows uninstaller. It deleted the script it was running, so it ended with "The batch file cannot be found"
  and exit code 1. That script is now removed last.
• --reinstall stopped at once with "A virtual environment already exists" on current versions of uv. It now
  rebuilds the Python environment again.

Every change is now installed for real on Windows, macOS and Linux by GitHub Actions before it is merged.

Already installed 1.0.0?

On an affected system, run the 1.0.1 installer once with --reinstall to replace PyTorch. Your models stay where they
are.

• Windows: in a Command Prompt in the unzipped folder, "Install on Windows.cmd" --reinstall
• Linux: sh "Install on Linux.sh" --reinstall

Download

System	Package	Then
Windows 10/11	AIRemove-1.0.1-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.0.1-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.0.1-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.0.1-all-platforms.zip

Then restart LightCraft once, select a raw photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac).

The models have their own licences (SAM License, CreativeML Open RAIL++-M, FLUX.1 [dev] Non-Commercial); see
THIRD-PARTY-NOTICES.md.

1.0.0

First release of AI Remove for LightCraft: remove objects and people from RAW photos with local AI, from inside
LightCraft. The result is written back into a RAW DNG stacked on the
original. It is an unofficial plug-in, not made or endorsed by the ArtCraft team.

Download

System	Package	Then
Windows 10/11	AIRemove-1.0.0-windows.zip	unzip, double-click Install on Windows.cmd
macOS 12+ (Apple Silicon)	AIRemove-1.0.0-macos.zip	unzip; in Terminal type sh , drag in Install on Mac.command, press Return
Linux x86-64	AIRemove-1.0.0-linux.tar.gz	unpack, sh "Install on Linux.sh"
All three	AIRemove-1.0.0-all-platforms.zip

The installer downloads Python, PyTorch for your GPU and the AI models you pick (5 to 30 GB). It then sets LightCraft
up. Restart LightCraft once afterwards, select a raw photo and press Ctrl+Shift+X (Cmd+Shift+X on a Mac).

What's in it

• One-click selection with SAM 3, refined with clicks, a brush and an eraser.
• LaMa, SDXL Inpainting and FLUX.1 Fill [dev] + Turbo variants, blended seamlessly.
• Writes a real RAW DNG: the fill is re-mosaicked into the sensor data with matching noise.
• Editable removals kept in a project file next to the raw.
• Windows, macOS and Linux, on NVIDIA, AMD, Intel Arc and Apple GPUs.
• English and Italian.
• An MCP server for AI assistants.

Tested

The Windows build is tested end to end on an AMD Radeon RX 7800 XT: install, re-run, window, MCP and uninstall. The
macOS and Linux installers are syntax-checked and their setup steps simulated, but not yet run on real machines:
reports are welcome.

The models have their own licences (SAM License, CreativeML Open RAIL++-M, FLUX.1 [dev] Non-Commercial); see
THIRD-PARTY-NOTICES.md.