A cross-platform, GPU-first raster image editor with local AI.

---

Disclaimer

Sable is provided as is and is under active development. There is no guarantee
of ongoing maintenance. You are free to open pull requests or fork it.

---

What is Sable?

Sable is a Photoshop/Affinity-style raster image editor that runs on Windows,
macOS, and Linux. It is GPU-first: the canvas is a real GPU surface (wgpu →
DX12 / Vulkan / Metal through one API), and the document is a non-destructive
node graph that the GPU compositor recomputes on every change — adjustments,
live filters, masks, and layer effects are all first-class, serializable,
undoable graph nodes, never baked-in pixels.

It also ships a local AI light tier that runs in-process on the GPU with no
Python and no cloud: smart selection, background removal, upscale, and object
removal. A heavier opt-in generative tier — Generative Fill and
text-to-image through a locally managed ComfyUI runtime — is fully decoupled:
nothing is installed until you enable it, and the base app stays lean and
offline.

Sable is MIT-licensed. AI model weights are always user-provided — Sable
ships pointers and shows each model's licence before download, never the
weights themselves.

Features

Editing

• GPU compositor with the full Photoshop + Affinity blend-mode set; linear-float
  working space.
• Non-destructive layer graph: pixel / shape / text / vector-path / group
  layers, adjustment layers, live filter layers, per-layer masks and effects.
• Adjustments — Brightness/Contrast, Levels (+ output levels), Curves, HSL,
  Exposure, Vibrance, Threshold, Posterise, Invert, B&W, White Balance, Colour
  Balance, Channel Mixer.
• Live filters — Gaussian / Box / Motion / Zoom blur, Sharpen, Unsharp,
  High Pass, Clarity, Add Noise, Denoise.
• Layer effects — drop shadow, outer/inner glow, stroke, colour & gradient
  overlay, bevel/emboss.
• Tiled (256×256) GPU-resident layer storage with a shared atlas + LRU residency
  for bounded VRAM regardless of document size.

Tools

• Brush / pencil / eraser (hardness, flow, HUD size-adjust), bucket fill,
  eyedropper, clone / heal / spot-heal / patch, liquify, mesh warp.
• Selections — marquee (rect/ellipse), lasso, magic wand, colour range, with
  add / subtract / intersect combine modes.
• Move + full affine transform (rotate, non-uniform scale, shear,
  perspective/free-distort), align & distribute.
• Vector pen / node tools and shapes (rectangle, rounded-rect, ellipse,
  line, polygon, star, arrow) with caps/joins/dashes; type with area text,
  text-on-path, and text-to-curves.

Local AI (light tier — no Python)

• Smart selection (SAM 2) — hover-to-select with automatic mask generation.
• Background removal (BiRefNet / RMBG).
• Upscale (Real-ESRGAN) with seamless tiling.
• Object removal / inpaint (LaMa).
• Built-in model manager: curated download pointers (licence shown per model),
  any-URL / HuggingFace import, per-task defaults, and VRAM-fit badges.

Generative AI (opt-in — ComfyUI)

• Generative Fill and Generate Image — fill a selection or create an
  image from a text prompt, driven by a local ComfyUI runtime that Sable
  installs and manages in the background. Nothing is installed until you
  enable it in Preferences.
• Reuses an existing ComfyUI install if one is found — checkpoints and
  LoRAs are referenced in place, never copied.
• Full workflow import — bring your own workflow, exported in ComfyUI's
  API format. Your graph stays the source of truth; Sable injects the image,
  prompt, seed, steps, CFG, LoRAs and model picks at run time, so any
  architecture your workflow supports just works.
• Generative panel with model presets, compatible LoRAs, prompt / negative
  prompt, steps, CFG, denoise, seed, and CPU offload for models larger than
  your VRAM.

Workspace

• Multi-document tabs, history panel with non-linear undo + named snapshots,
  command palette, rebindable hotkeys, light / gray / dark themes, rulers,
  guides, grid, and a customisable canvas overlay.
• Native .sable format (layers + history), plus PNG import/export.

Install

Pre-built installers are produced per release for all three platforms:

• Windows — sable-windows-x64-setup.exe (Inno Setup; associates .sable).
• macOS — sable-macos-x64.dmg / sable-macos-arm64.dmg (drag to
  Applications). The bundle is ad-hoc signed; on first launch use
  right-click → Open to bypass Gatekeeper for an unsigned build.
• Linux — Sable-x86_64.AppImage (chmod +x, then run).

Grab them from the Releases page.

Build from source

Requires the .NET 10 SDK. A GPU with DX12 / Vulkan / Metal is required to
run the app (the canvas is GPU-only); unit tests are headless.

dotnet build                                   # whole solution (Sable.slnx)
dotnet test  tests/Sable.Tests                 # headless unit tests
dotnet run   --project src/Sable.App           # run the editor
dotnet run   --project src/Sable.Gpu.Spike     # GPU smoke (compute + compositor -> PNG)

Packaging a release

Tag a commit (vX.Y.Z) and push — the
release workflow publishes a self-contained
single-file build per RID and packages it: Windows installer (Inno Setup),
macOS .app + DMG, and Linux AppImage. The app version is derived from the tag.

Architecture

Module	Responsibility
Sable.App	Avalonia shell (chrome, menus, panels, dialogs).
Sable.UI	MVVM view-models (CommunityToolkit.Mvvm).
Sable.Canvas	GPU surface host embedded in Avalonia; platform input seam.
Sable.Engine	Document graph, layers, GPU compositor.
Sable.Gpu	wgpu binding (Silk.NET.WebGPU) + WGSL shaders.
Sable.Imaging	Codecs, colour, tiling, IO.
Sable.Tools	Tool logic (brush, selection, transform, ...).
Sable.Ai	ONNX Runtime light tier + orchestration.
Sable.Format	.sable container IO + history.
Sable.Core	Math, colour, undo, settings, AI contracts.

The engine is UI-agnostic and headless-testable; the only per-OS code is the
canvas surface + input backend (IPlatformBackend / IInputSource).

Support the Project

If you find Sable useful and want to support its development:

Buy me a coffee on Ko-fi

License

Sable is released under the MIT License. Third-party components and
their licenses are listed in THIRD-PARTY-NOTICES.md.
AI model weights are not bundled and remain under their own licenses.