How do you install the Krita AI Diffusion plugin on Windows 10?
Install Krita 5.2.0 or newer, download the plugin ZIP from GitHub, then in Krita go to Tools → Scripts → Import Python Plugin from File, select the ZIP, and restart. Enable it via Settings → Dockers → AI Image Generation, then click Configure to auto-install the local server. A GPU with 6GB+ VRAM is strongly recommended.
Why — the first-principles explanation
The install trips people up because it's really two installs wearing one coat, and the instructions rarely say so.
The first is trivial: the plugin itself. It's a small Python package that adds a panel to Krita. Krita has a built-in importer for exactly this, which is why the step is one menu item. Nothing heavy happens here.
The second is the part that actually takes time and disk space, and it isn't the plugin. The plugin does not generate images. It's a front end that talks to ComfyUI, a separate image-generation backend, which in turn needs the actual diffusion model weights — multi-gigabyte files. When you click Configure, the plugin downloads and sets up that whole stack for you. That's why the install feels enormous relative to a ZIP file: you're pulling a server and its models, not an add-on.
Understanding that split explains every choice you'll face. Why 6GB of VRAM? Because diffusion runs iteratively — it starts with noise and denoises across many steps, and the entire model has to sit in graphics memory the whole time. Below roughly 6GB you're either swapping constantly or out of memory. CPU processing works and is, per the official docs, "very slow" — that's not a caveat, it's the honest description, because a GPU does thousands of these operations in parallel and a CPU does a handful.
Why is there a cloud option? Because the backend is separable. If the plugin merely talks to ComfyUI, that ComfyUI doesn't have to be on your machine. Same plugin, someone else's GPU. That's the escape hatch when your hardware can't clear the bar.
Why does it connect automatically sometimes? Because if ComfyUI is already running locally when you launch Krita, the plugin will automatically try to connect to it — and the same mechanism serves remote servers. Once you know it's a client looking for a server, the behavior stops being mysterious.
On Windows 10 specifically, the plugin supports NVIDIA via CUDA and AMD via ROCm, plus Intel via XPU. NVIDIA is the best-supported path. Nothing here is Windows 10-exclusive — Windows, Linux, and macOS all work, and the managed install handles the platform details.
An example that makes it click
Think of it like installing a TV remote versus installing the TV.
The remote is small. It arrives in a padded envelope, you pop the batteries in, done — two minutes. That's the Krita plugin: one menu item, a ZIP file, a restart.
But a remote doesn't show anything. Somebody still has to bring in the actual television, bolt it to the wall, and run the power. That's ComfyUI and the model weights — several gigabytes, and the reason the "install" takes far longer than a ZIP should. When you press Configure, you're not setting up the remote. You're ordering the TV.
And the 6GB VRAM rule is just the wall having to hold the weight. You can technically prop the TV on a folding chair — that's CPU mode — and the docs will politely tell you it's very slow.
How to do it
- Install or update Krita. The plugin requires version 5.2.0 or newer — check Help → About Krita if you already have it.
- Check your GPU. A graphics card with at least 6 GB of VRAM is strongly recommended for local generation. NVIDIA (CUDA) is the best-supported path on Windows 10; AMD (ROCm) and Intel (XPU) are also supported.
- Download the latest plugin release ZIP from the official GitHub repository (github.com/Acly/krita-ai-diffusion). Do not unzip it.
- In Krita, go to Tools → Scripts → Import Python Plugin from File and select the downloaded ZIP archive.
- Restart Krita after the installation completes. The plugin won't appear until you do.
- Enable the panel: Settings → Dockers → AI Image Generation. The docker appears in your workspace.
- Click Configure in the plugin docker to start the automatic local server setup. This downloads and installs ComfyUI plus the diffusion model weights — expect several gigabytes and a long wait.
- Alternative if your hardware is too weak: use the cloud generation option, or run ComfyUI yourself before launching Krita — the plugin will automatically try to connect to an already-running local or remote server.
Key facts
- The Krita AI Diffusion plugin requires Krita 5.2.0 or newer.
- Installation is via Krita's built-in importer: Tools → Scripts → Import Python Plugin from File, selecting the downloaded ZIP, followed by a restart.
- A graphics card with at least 6 GB of VRAM is strongly recommended for local generation.
- The plugin uses ComfyUI as its backend — it is a front end, not the image generator itself; clicking 'Configure' in the docker triggers automatic local server setup and model downloading.
- Supported hardware includes NVIDIA via CUDA (Windows/Linux), AMD via ROCm (Windows/Linux), Intel via XPU (Windows/Linux), and Apple Silicon via MPS (macOS 14+).
- CPU processing is supported but officially described as 'very slow.'
▶ The 60-second explainer (script)
How do you install the Krita AI Diffusion plugin on Windows 10? Here's the thing that makes this confusing: it's two installs wearing one coat. Install one is trivial. Make sure you have Krita 5.2.0 or newer. Download the plugin ZIP from the official GitHub repo — don't unzip it. In Krita, go to Tools, Scripts, Import Python Plugin from File, pick the ZIP. Restart Krita. Then turn the panel on at Settings, Dockers, AI Image Generation. That's it. Two minutes. But now install two — and this is the part nobody warns you about. The plugin doesn't generate images. It's a front end. It talks to ComfyUI, a separate backend, which needs the actual diffusion model weights — multi-gigabyte files. So when you click Configure in the docker, it downloads and sets up that entire stack. That's why an install that started with a small ZIP suddenly takes forever. It's like a TV remote versus a TV. The remote arrives in a padded envelope, batteries in, done. But a remote doesn't show you anything — somebody still has to bring in the television and bolt it to the wall. When you press Configure, you're not setting up the remote. You're ordering the TV. That split explains everything else. Why 6 gigabytes of VRAM? Because diffusion runs iteratively — it starts with noise and denoises over many steps, and the whole model has to stay in graphics memory the entire time. Below about 6GB you're swapping or you're out of memory. CPU mode works, and the official docs call it, quote, very slow. That's not a caveat — a GPU does thousands of these operations in parallel, a CPU does a handful. Why is there a cloud option? Because the backend is separable. If the plugin just talks to ComfyUI, that ComfyUI doesn't have to be on your machine. Same plugin, someone else's GPU. And if it ever connects on its own — that's because if ComfyUI is already running when Krita launches, the plugin tries to connect automatically. On Windows 10: NVIDIA via CUDA is the best-supported path, AMD ROCm and Intel XPU also work.
What authoritative sources say
People also ask
Why is the install so large if the plugin is a small ZIP?
Because the ZIP is only the front end. The plugin doesn't generate images — it talks to ComfyUI, and clicking Configure downloads that backend plus multi-gigabyte diffusion model weights.
Can I run it without a good GPU?
Yes, two ways. CPU processing is supported but officially described as 'very slow.' The better option is cloud generation, since the backend is separable from the plugin — same panel, someone else's GPU.
Which Krita version do I need?
5.2.0 or newer. Check Help → About Krita. Older versions won't load the plugin, and this is the most common reason the docker never appears.
The AI Image Generation docker isn't showing up. What's wrong?
Almost always one of two things: you didn't restart Krita after importing, or you're on a Krita version below 5.2.0. Restart first, then check Settings → Dockers → AI Image Generation.
Does AMD work on Windows 10?
Yes — AMD is supported via ROCm on Windows and Linux, as is Intel via XPU. NVIDIA through CUDA remains the best-supported path.