Paint an area. Only that area changes.
Brush over a part of the photo, then choose Change, Remove, Replace or Add. The model sees exactly where you painted, and every pixel outside the brush stays as it was.
BetaQwen-Image 2.1 Turbo, on your own computer
Edit photos by describing the change.
Kokon runs Qwen-Image 2.1, the best open image-editing model, on your own GPU. One installer for Windows and Mac. No Python, no ComfyUI, no cloud, no account.
Free and open source · Windows 10/11 with a 12 GB GPU · Apple Silicon Macs
Runs on
- NVIDIA
- AMD
- Intel Arc
- Apple Silicon
- Windows
Examples
One sentence. Any edit.
Drag the slider. Every image on this page was made by Kokon on a single RTX 5070 Ti, first try.
Times on an RTX 5070 Ti: 22 to 46 s for a new image, 25 to 55 s for an edit. The prompts and seeds are in the repository.
Features
Made for people, not for pipelines.
No nodes, no samplers, no VRAM math. Write what you want; Kokon handles the rest.
Zero settings.
Size, memory and upscaling are chosen for your GPU. Pick Fast, Balanced or Quality. That's it.
FastBalancedQuality
Every version kept.
Each edit becomes a new version. Compare before and after, undo, or branch from any step.
Private by design.
Your photos and prompts never leave your computer. Turn on private mode and nothing is kept at all: no gallery, no files, no history.
0images uploaded, ever
No Python. One installer.
A 35 MB native app. The models download and are checked on first run, with resume.
35MB installer
Works with your AI agent.
A built-in MCP server and a local HTTP API let Claude Code, Cursor or any MCP client create and edit images on your machine, for free and offline.
generate_imageedit_imageinpaint_imageoutpaint_imageupscale_imageHow it works
Run Qwen-Image 2.1 locally in three steps.
No Python environment, no ComfyUI workflows, no model files to hunt down. Read the full guide.
Download the installer
A 35 MB native app for Windows 10/11 or macOS on Apple Silicon. It installs for your user, no admin rights needed.
Let it set up the model
Kokon detects your GPU (NVIDIA, AMD, Intel or Apple Silicon), picks the right preset and downloads Qwen-Image 2.1 Turbo, with resume and SHA-256 checks.
Describe the change
Drop a photo, write what you want and press Enter. Or paint an area first, and only that area changes.
Compare
The simple way to run Qwen-Image.
Node-based tools give you every knob; cloud editors need an account and an upload. Kokon is for people who want the result, on their own machine.
| Kokon | Node-based tools e.g. ComfyUI | Cloud editors | |
|---|---|---|---|
| Setup | One installer | Python environment, custom nodes, workflows | Account |
| Your photos stay on your computer | Yes | Yes | No, they are uploaded |
| Cost | Free | Free | Credits or subscription |
| Edit with one sentence | Yes | After building a workflow | Yes |
| Paint an area, change only that | Built in | With an inpainting workflow | Some |
| Settings to tune | None (Fast, Balanced, Quality) | Many | Few |
| NVIDIA, AMD, Intel, Apple Silicon | Yes (AMD, Intel, Mac in beta) | Depends on the setup | Any browser |
| Local MCP server for AI agents | Built in, no API key | Third-party | Paid API |
AI agents
Your agent can make images now.
Register Kokon once. Then ask for a hero image, a cleaned-up screenshot or a product shot, and your agent makes it on your GPU. The app doesn't even need to be open.
claude mcp add --scope user kokon -- "$env:LOCALAPPDATA\Kokon\kokon-mcp.exe"Windows command shown. Mac and other hosts: MCP setup guide.
my-website
> Make a hero image for my landing page: a blue morpho butterfly on a silk cocoon, 16:9
● kokon · generate_image
aspect_ratio: "16:9" · quality: "quality"
⎿ Saved public/hero.png · 2752×1536 · 39 s
Requirements
Runs on the computer you have.
Windows
SystemWindows 10 or 11, 64-bit
NVIDIA12 GB VRAM or more (RTX 3060 12 GB, 4070, 5070…)
AMD, Intel Arc12 GB or more, through VulkanBETA
DiskAbout 15 GB for the models
macOSBETA
SystemmacOS 26 or newer
ChipApple M1 or newer (Pro, Max or Ultra recommended)
Memory32 GB recommended · 16 GB works, more slowly
DiskAbout 15 GB for the models
Privacy
Nothing leaves your computer.
Creating, editing, the gallery, the MCP server and the API all talk to 127.0.0.1 only. Kokon goes online for two things, and you can check it yourself.
Setup
The engine from github.com and the models from huggingface.co, checked against pinned SHA-256 hashes.
Update check
At most once a day, a small file with the newest version number. Nothing about you is sent, and you can turn it off.
Private mode
One click and nothing is saved: no gallery, no files, no project history.
FAQ
Questions.
Is it really free?
Yes. Kokon is open source under the GNU AGPL-3.0 license. No account, no credits, no subscription. It uses your own GPU, so there is nothing to pay for.
How do I run Qwen-Image 2.1 locally without ComfyUI?
Install Kokon and let its setup download the model. Kokon runs Qwen-Image 2.1 through stable-diffusion.cpp, a native engine, so there is no Python, no ComfyUI and no workflow to build. The step-by-step guide is at kokon.app/qwen-image-2-1-local.
Which model does it use?
Qwen-Image 2.1 Turbo, the official 8-step version of Qwen's image model, which both creates and edits images and leads the open-weights image-editing rankings.
Does it work on AMD or Intel GPUs?
Yes, in beta, through the Vulkan engine on Windows with 12 GB of VRAM or more. On NVIDIA it uses CUDA; on Apple Silicon Macs it uses Metal.
My GPU has 8 GB. Will it work?
Not well today. 12 GB is the smallest size that gives good results with this model. On a Mac, 16 GB of unified memory works but slowly; 32 GB is recommended.
Can Claude Code or Cursor generate images with it?
Yes. Kokon ships an MCP server: register it once and any MCP client can create, edit, inpaint, expand and upscale images on your GPU, for free and without an API key.
Is it an alternative to Photoshop Generative Fill?
For many edits, yes: paint an area and describe what should be there, or remove, replace and add objects. It runs offline and costs nothing, but it is an AI editor, not a full layer-based image editor.
Can I use the images commercially?
Not with the default model. The app is open source (AGPL-3.0), but Qwen-Image 2.1 is released under the Qwen Research License: research and evaluation only, no commercial use without a license from the Qwen team. The optional upscaler is also non-commercial (CC BY-NC-SA 4.0).
Windows says the app is unrecognized. Is it safe?
The beta installer is not code-signed yet, so SmartScreen warns about it. Click "More info" and then "Run anyway". The code is public and every download the app makes is checked against a pinned hash.
Does it work offline?
Yes. After setup has downloaded the models, everything runs offline.
Describe it. Kokon makes it.
Free, open source, and yours to run.
Made by Gabriele LozuponeAI researcher. Built Kokon with the help of Claude Code. More about me →