Krita AI Diffusion

Krita AI Diffusion adds generative image tools to Krita's painting canvas. You can select part of an illustration or photo to fill, extend an image beyond its edges, refine existing content, or generate a new starting point without moving your work to a separate image generator. Its main appeal is that brushwork, layers, and selections remain part of the process: you can draw a rough idea, guide the model, and keep editing in Krita.

Paint first, then guide generation

Selections focus a change on the part of an image you want to work on. That makes the plugin useful for replacing an object, expanding a scene, or adding detail to a small area while keeping the rest of the canvas as context. A strength control lets you decide how closely a new result should follow the existing image. For broader changes, supported edit models can respond to text instructions.

Control layers provide more specific guidance. A sketch, line drawing, pose, depth map, or reference image can steer a result in ways a text prompt alone may not. Regions attach prompts and references to areas defined by Krita layers, so different subjects can have their own instructions. The project's guide cautions that regions constrain prompts; they do not guarantee an exact composition or shape. A sketch or other control layer is better suited to that job.

Live painting interprets changes to the canvas as you work, giving you a way to explore an idea before committing to a full generation. The plugin also supports upscaling, a job queue, generation history, and presets for models and settings. Available controls and results depend on the selected model and backend.

Local or online generation

Krita AI Diffusion uses ComfyUI behind the scenes. Its managed local server can install the backend and recommended model workloads, or you can connect a compatible ComfyUI server you already run. A custom server gives you more control, but you must provide the required models and custom nodes yourself. The separate online service is an option if you do not want to set up local GPU inference; it requires an account.

Using it on Linux

Install Krita first, then import the plugin's release ZIP through Krita's Python plugin menu and enable its AI Image Generation docker. For the managed server on Linux, Krita's AppImage is the recommended installation path as some packages cannot be installed from the Flatpak build.

Local generation needs a suitable backend and model files. Depending on your hardware, local generation on Linux can use CUDA with NVIDIA, ROCm with AMD, or XPU with Intel GPUs. A CPU can also run it, but very slowly. Model downloads can take substantial disk space, and memory needs vary with the model and workload. An existing ComfyUI installation can avoid a second backend setup, provided its required extensions and models are present.

Strengths and things to consider

The close connection to Krita is the reason to choose this plugin. Selections, painted guides, layer-based regions, and iterative refinement let you use generative tools within an image you are already editing. You can keep a local setup or choose the online service when local hardware is unsuitable.

Setup is the main trade-off. The plugin is a separate project from Krita, the local server adds GPU and model dependencies, and a custom ComfyUI connection needs compatible nodes. Some techniques also require learning how prompts, layers, and control images interact. Krita AI Diffusion is best for artists who want AI assistance inside a painting workflow and are willing to check backend compatibility before installing.