Kokon

Why Kokon

Examples

Every image below was made by Kokon itself (no stock photos), first try with that seed. Prompts, seeds and times are in marketing/showcase/PROMPTS.md.

Install

1. Download from the latest release:

• Windows: Kokon_<version>_x64-setup.exe. Run it. It installs for your user only, so you don't need administrator rights.

• Mac (Apple Silicon, beta): Kokon_<version>_aarch64.dmg. Open it and drag Kokon to Applications. The first time, macOS blocks it because the app is not notarized by Apple: open System Settings → Privacy & Security and click Open Anyway, or run one Terminal command if there is no such button (details).

• Linux (x86-64, beta): Kokon_<version>_amd64.deb for Ubuntu, Debian and Mint, or Kokon-<version>-1.x86_64.rpm for Fedora. Install it with your package manager, which also pulls in the few libraries it needs:

  sudo apt install ./Kokon_<version>_amd64.deb      # Ubuntu 22.04+, Debian 12+, Mint 21+
  sudo dnf install ./Kokon-<version>-1.x86_64.rpm   # Fedora 40+
  

  For NVIDIA cards you need the proprietary driver (the CUDA runtime comes with Kokon). AMD and Intel cards use Mesa's Vulkan driver, installed on most desktops already. If apt prints "Download is performed unsandboxed as root", that's harmless (apt can't read your home folder); copy the .deb to /tmp first to avoid it.

2. Set up models. On first launch Kokon shows your graphics card (or your Mac's chip and memory), recommends a preset and downloads everything it needs. The download resumes if it is interrupted and every file is checked against its SHA-256 checksum. On a fast connection the whole thing takes a few minutes, and the setup screen shows what Kokon can do while you wait.

3. Edit. Drop or paste a photo, describe the change and press Enter.

Requirements

Windows

Mac (beta)

Linux (beta)

Kokon measures your machine's speed with the first image and shows realistic time estimates from then on.

[!NOTE] NVIDIA cards run on the CUDA engine, the fastest and tested path. AMD and Intel cards use the Vulkan engine (beta). First user report on AMD: an RX 9060 XT 16 GB on Windows made a 16:9 1080p image at Balanced in about 80 s (the first image after start, model loading included), with edits quicker after that. Intel is still untested, so please report how it goes. Vulkan is slower than CUDA, and its first image takes longer. Macs use the Metal engine (beta), tested so far on a 16 GB M2 Pro. On Linux Kokon ships its own CUDA and Vulkan builds of the same engine, tested so far on an RTX 5070 Ti under WSL. Details in docs/platforms.md and docs/linux.md.

The setup screen picks a preset from your VRAM. On a Mac the presets are named by unified memory (16–24 GB Mac, 32–36 GB Mac, 48 GB Mac, 64 GB+ Mac), because the graphics part can use only about two thirds of it and the text encoder shares the same memory. You can change the preset later from Settings → Models → Set up models…

Every preset also gets stable-diffusion.cpp (the CUDA 12 build on NVIDIA, the Vulkan build on AMD and Intel, the Metal build on a Mac), the Qwen-Image 2.1 VAE with madebyollin's texture fix (it removes the faint 2-pixel checkerboard of the original decoder; installs from 0.9.2 or earlier switch to it when you accept the update), the Qwen3-VL vision projector and, on a PC, the Superscale-SP upscaler (67 MB). The image model is Qwen's official 8-step Qwen-Image 2.1 Turbo through DogukanUrker/Qwen-Image-2.1-Turbo-GGUF; the other sources are Qwen/Qwen3-VL-8B-Instruct-GGUF, madebyollin/texture-fix-vae-for-qwen-image-2.1 and Comfy-Org/Qwen-Image-2.1 (the original VAE).

Files go to %LOCALAPPDATA%\KokonData\sdcpp (Mac: ~/Library/Application Support/KokonData/sdcpp) by default. You can point the setup screen at an existing stable-diffusion.cpp folder instead; files that are already there are kept.

Already have Qwen-Image 2.1 files, for example in ComfyUI? Choose Use models I already have in the setup screen and pick the folder (such as ComfyUI\models): Kokon finds the image model, VAE, text encoder and vision projector by their contents, as GGUF or safetensors in any quantization, and downloads only the engine and any piece you don't have. Files stay where they are. Your own files are not tested by Kokon; you can switch to the official files at any time in Settings → Models.

What you can do

New to prompts? The prompt guide shows good and bad prompts side by side, with real results. Every result can be edited again, varied, used as a reference, upscaled, saved, copied or opened in Explorer, Finder or your file manager. A few things make that loop fast:

• Before/after: compare with Slider, Flicker or Changes, or hold B to see the original.
• Undo/redo between versions with Ctrl+Z / Ctrl+Y (⌘ on a Mac). Undo returns to the version the current one came from, and nothing is ever deleted.
• Gallery: every image keeps its full recipe (prompt, seed, settings, pipeline) and reopens with one click. The recipe is also embedded in the PNG.
• Quality presets: Fast (the default, ~1 MP), Balanced (~1.5K) and Quality (native 2K). All run Qwen's official 8-step Turbo model with its own schedule, so you never need to tune steps or guidance.

How it works

flowchart LR
    A["📷 Your photo<br>+ a sentence"] --> B["Planner<br><sub>mode · size · memory budget</sub>"]
    B --> C["Qwen-Image 2.1<br><sub>stable-diffusion.cpp<br>on your GPU</sub>"]
    C --> D["AI upscaler<br><sub>Superscale-SP, back to<br>full size</sub>"]
    D --> E["🖼️ Result<br><sub>gallery + recipe</sub>"]
    M["🤖 Claude · Cursor · scripts"] -. "MCP / local API" .-> B

The planner decides what kind of job your request is, picks a working size that fits your GPU and runs a single diffusion pass. When the result has to be larger than what the model rendered, it is enlarged with an AI upscaler instead of a plain resize. For brushed edits, Kokon marks the area on a second reference image so the model knows exactly where to work. Without that, Qwen-Image 2.1 cannot see the mask.

Want the details? The engineering notes cover the architecture, the measured backend settings and the reasons behind each default.

Use it from your AI agent

Kokon ships kokon-mcp, an MCP server. Register it once with Claude Code:

# Windows
claude mcp add --scope user kokon -- "$env:LOCALAPPDATA\Kokon\kokon-mcp.exe"

# Mac
claude mcp add --scope user kokon -- /Applications/Kokon.app/Contents/MacOS/kokon-mcp

Or, on either system with Node.js 22+, through npm (still needs the app installed):

claude mcp add --scope user kokon -- npx -y kokon-mcp

Then ask: “Generate a hero image for this README and save it in docs/” or “Remove the background clutter from screenshot.png”. The app doesn't need to be open, because the MCP server starts the engine on demand. The first call takes about a minute while the model loads.

Tools: generate_image · edit_image · edit_images · inpaint_image · outpaint_image · upscale_image · get_image_info · get_generator_status

Project-scoped .mcp.json, or any host that takes the standard format (on a Mac the command is /Applications/Kokon.app/Contents/MacOS/kokon-mcp):

{
  "mcpServers": {
    "kokon": {
      "type": "stdio",
      "command": "C:\\Users\\<you>\\AppData\\Local\\Kokon\\kokon-mcp.exe",
      "args": []
    }
  }
}

Same thing from the command line:

claude mcp add-json --scope user kokon '{"type":"stdio","command":"C:\\Users\\<you>\\AppData\\Local\\Kokon\\kokon-mcp.exe","args":[]}'

• Every result returns the absolute path of the saved file, its size, seed and generation metadata.
• Paths may be absolute or relative to the project (CLAUDE_PROJECT_DIR). Reads and writes are limited to the project folder, Kokon's output folder and folders listed in QLS_MCP_ALLOWED_DIRS (separated by semicolons). Existing files are never overwritten.
• Images are not sent back to the model by default; they stay on disk. Pass include_preview: true to include a small JPEG preview.
• upscale_image takes scale (default 2) or width/height (up to 8192 px).
• Progress is reported through MCP progress notifications. Claude Code's default tool timeout is 30 minutes (MCP_TOOL_TIMEOUT).
• Without the desktop window open, the model is unloaded after 15 idle minutes (Settings → General).

Base URL http://127.0.0.1:39151/api/v1. Every call except /ping needs Authorization: Bearer <token>, with the token from %LOCALAPPDATA%\KokonData\service.json (Mac: ~/Library/Application Support/KokonData/service.json). Requests that carry a browser Origin header are refused.

$i = Get-Content "$env:LOCALAPPDATA\KokonData\service.json" | ConvertFrom-Json
$h = @{ Authorization = "Bearer $($i.token)" }
Invoke-RestMethod -Method Post "http://127.0.0.1:$($i.port)/api/v1/generate" -Headers $h `
  -ContentType 'application/json' -Body '{"prompt":"a lighthouse at dusk, 16:9","quality":"fast"}'

Simple-operation fields: prompt, image / images[], mask, action (edit|remove|replace|add, inpainting), width, height, aspect_ratio, quality (fast|balanced|quality), seed, output_path, preserve_dimensions, private, left/right/top/bottom + units (pixels|percent) for outpainting, scale for upscaling. Full requests (/jobs) also accept "mode": "upscale" with "upscale": {"scale" | "width"/"height"}.

Privacy

Kokon makes only two kinds of network request:

• Setup: github.com for the stable-diffusion.cpp release (and the MoSR upscaler, if the copy that comes with the app is missing) and huggingface.co for the models and the other upscalers.
• Update check: at most once a day, the app fetches latest.json from this repository's latest GitHub release, a small file with the newest version number. Nothing about you or your images is sent. Turn it off on the welcome screen or in Settings → General. Updates are downloaded only when you click Update, and installed only if their signature matches the key built into the app.

Creating, editing, the gallery, the MCP server and the API talk to 127.0.0.1 only. Logs never contain your prompts.

You can check this yourself. While you work, watch Kokon, kokon-service and sd-server in Resource Monitor's Network tab on Windows (or run Get-NetTCPConnection -OwningProcess <pid>), or with lsof -i -P | grep -iE "kokon|sd-server" on a Mac.

Private mode (the eye icon in the top bar) keeps images and prompts out of the gallery, the output folder and the project list. They are deleted when you turn private mode off, close the window, or restart the service. Save as… still writes a copy wherever you choose.

How it compares

Kokon does one thing: editing photos with Qwen-Image 2.1, with as little setup as possible. If you want full control over every node, ComfyUI is the better tool.

FAQ

SmartScreen shows this for new apps that don't have a download history yet. Click More info → Run anyway. You can check the installer against the SHA-256 hash listed in the release notes, or build it yourself.

Kokon is not notarized by Apple (that needs a paid Apple developer account), so macOS blocks it the first time. You only do this once; in-app updates don't trigger it again.

1. Open Kokon. macOS says “Kokon” Not Opened: click Done (not Move to Trash).

2. Open System Settings → Privacy & Security and scroll down to Security. Next to “Kokon” was blocked to protect your Mac, click Open Anyway, then confirm with your password or Touch ID.

3. Open Kokon again and click Open in the last prompt.

If macOS instead only says The application “Kokon” can't be opened, with just an OK button, and Privacy & Security shows no Open Anyway: open Terminal and run

xattr -dr com.apple.quarantine /Applications/Kokon.app

then open Kokon as usual. This removes the "downloaded from the internet" flag from Kokon only; Gatekeeper stays on for everything else.

You can also build it yourself.

Only to download the models. After that Kokon works fully offline. If you are online it checks once a day for a new version, which you can turn off on the welcome screen or in Settings → General.

AMD and Intel: in beta. The setup screen installs stable-diffusion.cpp's Vulkan build for them, which uses the graphics driver you already have. A user's RX 9060 XT 16 GB made a 1080p image at Balanced in about 80 s (first image, model loading included); Intel Arc is still untested, and integrated graphics are not supported. Please open an issue with your card and how it went.

Mac: yes, in beta, on Apple Silicon (M1 or newer) with macOS 15 Sequoia or newer. It runs on a 16 GB Mac in a lighter mode, but slowly (about 4 minutes per 1-megapixel image on an M2 Pro); 32 GB or more and a Pro, Max or Ultra chip are recommended. See the requirements and docs/platforms.md.

Linux: yes, in beta, on x86-64 with a .deb (Ubuntu, Debian, Mint) or an .rpm (Fedora). NVIDIA cards run Kokon's own CUDA build of the engine, AMD and Intel cards its Vulkan build. A headless server with the interface in your browser, and a Docker image, come next (#15).

Yes, in beta since 0.9.3. Setup picks the 8 GB preset: the same files as the 12 GB one, with smaller working sizes, so images are made at about 1 megapixel (then enlarged). Users report it works on an RTX 3060 Ti and on an RTX 5050 Laptop, where an edit takes 1–2 minutes. The text encoder runs from system RAM, so 16 GB of RAM or more is recommended. If you try it, please tell us how it runs in an issue.

Not with the default model. Kokon's code is open source (see License), but the image model has its own license: Qwen-Image 2.1 and its Turbo model are released under the Qwen Research License, which allows research and evaluation only, no commercial use. That includes selling or using commercially the images you make with it. For commercial use, ask the Qwen team for a license ([email protected]).

The Qwen3-VL text encoder is Apache-2.0. The default upscaler, Superscale-SP by Nmkd, is WTFPL, and the experimental MoSR by Philip Hofmann is CC BY 4.0: both allow commercial use. The optional UltraSharp upscaler is CC BY-NC-SA 4.0, non-commercial only.

HEIC/HEIF photos open like any other image. On a Mac they are decoded by macOS itself. On Windows, Microsoft's HEIF Image Extensions and HEVC Video Extensions must be installed.

iPhone photos with an HDR gain map (iOS 18+) show a Keep HDR switch in Advanced, which is on by default. Edits of such photos are saved as Ultra HDR JPEG: Chrome, Edge, Android and recent Apple systems show them in HDR, and every other viewer shows a normal JPEG. This needs ffmpeg 7.1+ on PATH; without it the result is saved in SDR and the job's notes say why.

By default Superscale-SP by Nmkd (ESRGAN), downloaded by the setup screen; on a Mac, Lanczos, because AI upscalers take minutes there (Superscale-SP: 5 min 45 s to enlarge a 1024×1024 image to 2048×2048 on an M2 Pro, against about a second). Pick another one in the Advanced menu next to the prompt, or in Settings → Models; Compare upscalers there shows each one on the same crops (a face, a JPEG photo, a night street, text, a landscape, a 4× upscale):

• Superscale-SP (default): sharp and natural on most images; on JPEG photos and dark areas it can draw a faint grid.
• MoSR (experimental, 17 MB, comes with the app): the fastest, good for 4× upscales, but on clean images enlarged less than about 3× it draws a fine grid. Kokon contributed it to stable-diffusion.cpp (pull request #2128, merged 2026-10-12).
• UltraSharp (optional download): very clean, but smooths skin into an artificial look and can invent lines. Non-commercial license.
• Lanczos: a plain resize, never invents anything.

Upscaler strength (0–100 %) mixes the upscaler's result with a plain resize, to tone down texture it invents. Settings → Models also has Shrink the image before upscaling, which gives the upscaler a smaller image, like the ones it was trained on.

Windows: Settings → Apps → Installed apps → Kokon → Uninstall. The uninstaller asks whether to delete the downloaded models and app data too (%LOCALAPPDATA%\KokonData).

Mac: drag Kokon from Applications to the Trash. To free the model space too, delete ~/Library/Application Support/KokonData.

Your images stay in your Pictures folder.

If a download fails, the setup screen shows the file's download page and target folder so you can fetch it by hand. For anything else, please open an issue with your GPU (or Mac model), driver or macOS version and the log from KokonData/logs (see the table above).

Roadmap

• Create, edit, combine, brush edits, expand, upscale
• First-run setup with GPU detection and verified, resumable downloads
• MCP server and local HTTP API
• Before/after, version undo, private mode, HEIC and HDR
• macOS on Apple Silicon (beta, Metal)
• One-click in-app updates on Windows, Mac and Linux
• Linux (.deb and .rpm, beta)
• Colour LUTs: built-in looks and your own .cube files
• Qwen's official 8-step Turbo model for every preset
• AMD and Intel GPUs through Vulkan (beta now; works on NVIDIA, to verify on AMD/Intel)
• Background removal and transparent cutouts
• Batch edits: one instruction for a whole folder
• LoRA support, with a curated list tested with Kokon (0.10.0)
• Headless server with the interface in the browser, then Docker (#15)
• Plugins for Krita, GIMP and Photoshop through the local API

Have an idea? Start a discussion or 👍 an existing one.

Build from source

Windows: Rust (stable, MSVC), Node 22+, Visual Studio Build Tools and WebView2.

npm install
cargo test -p qls-core            # core unit tests
npx vitest run                    # UI logic tests
cd mcp; npx vitest run; cd ..     # MCP tests (includes a stdio protocol test)
npx tauri dev                     # run the desktop app in development
powershell -ExecutionPolicy Bypass -File scripts\build-release.ps1   # installer

Mac (Apple Silicon, macOS 15+): Xcode Command Line Tools, Rust and Node 22+.

bash scripts/dev/mac-dev.sh --check   # check the prerequisites
bash scripts/dev/mac-dev.sh           # build and run the app in development
bash scripts/build-release-mac.sh     # Kokon.app + .dmg in dist-release/

Linux (x86-64): Rust (stable, via rustup), Node 22+ and Tauri's system packages. Build on the oldest distribution you want to support (releases are built on Ubuntu 22.04).

bash scripts/build-release-linux.sh --setup   # once: system packages
bash scripts/build-release-linux.sh           # .deb + .rpm in dist-release/

Publishing a release (signed in-app updates for every platform) is described in docs/releasing.md.

flowchart TB
    UI["<b>Kokon</b><br>Tauri 2 + React"]
    MCP["<b>kokon-mcp</b><br>MCP over stdio"]
    S["Scripts and other apps"]
    SVC["<b>kokon-service</b><br>planner · preprocessing · job queue · gallery<br><code>127.0.0.1:39151/api/v1</code>"]
    SD["<b>sd-server</b><br>stable-diffusion.cpp (CUDA · Vulkan · Metal)<br><code>127.0.0.1:39150</code>"]
    UI -- "HTTP + SSE, token" --> SVC
    MCP -- "HTTP, token" --> SVC
    S -- "HTTP, token" --> SVC
    SVC -- "native async API" --> SD

The Rust core (crates/qls-core) holds all the logic: geometry, image processing, planning, the inference adapter, the job queue and the gallery. The service, the desktop app and the MCP server are thin layers on top. See the engineering notes for a module map and the backend facts each default is based on.

Contributing

Bug reports, translations, example prompts and pull requests are all welcome. Issues tagged good first issue are a good place to start. Please run the tests above before opening a pull request.

Responsible use

Kokon edits real photos. Don't use it to deceive, harass or impersonate people, or to make images of someone without their consent. You are responsible for what you create and share.

License

Kokon's source code is released under the GNU Affero General Public License v3.0 (AGPL-3.0-only). You can use, study, change and share it freely. If you distribute a modified version, or offer one as a network service, you must publish its source under the same license. The license covers the code, not the images you make with it.

Want to ship Kokon inside a closed-source product or service? A commercial license is available: open a discussion or contact the author.

Model weights and the upscaler are downloaded from their authors and keep their own licenses (see the FAQ).

Kokon builds on stable-diffusion.cpp by leejet, Qwen-Image 2.1 and its Turbo model by the Qwen team, DogukanUrker's Turbo GGUF conversions, Tauri and the upscalers by Nmkd (Superscale-SP), Philip Hofmann (4xNomos2 MoSR) and Kim2091 (UltraSharp). Thank you.