Krita AI Diffusion brings targeted AI editing to the canvas

What this is about

Krita AI Diffusion, also called “Generative AI for Krita” by the project, adds generative image editing to the painting application Krita. Instead of ordering a complete image in a browser, users can select and modify areas on their existing canvas, then continue editing by hand. The plugin is developed in the Acly/krita-ai-diffusion repository; its associated product website is Interstice.

This tool check, dated November 12, 2026, describes an available tool, not a launch happening today. It is based on live checks of the project documentation and the cloud service’s statements. We did not perform an installation test or an image-quality comparison. The plugin is particularly relevant to illustrators, concept artists and people already working with layers, selection masks and brushes.

What Krita AI Diffusion actually does

The plugin can generatively fill selected image regions, add or remove objects and extend scenes beyond their previous boundaries. Users can control the strength of a revision. Reference images, sketches, line art and other control inputs guide the model. Regions allow different descriptions to be assigned to different image areas. A job queue and history help manage variations.

The key distinction is between the interface and computation: Krita remains the workspace, while the plugin connects it to a backend. According to the installation guide, users can choose a managed local server, their own ComfyUI installation or the online service. A custom ComfyUI instance needs compatible extensions and models; an arbitrary existing installation is therefore not automatically compatible.

To get started, users install the plugin ZIP through Krita’s Python plugin import. After restarting, the “AI Image Generation” docker becomes available. Version matching matters: the guide distinguishes Krita 5.x with plugin 1.x from an upcoming 2.x branch for Krita 6.x. Before downloading, check the currently specified combination rather than simply choosing the highest version number.

Why it matters

The practical benefit is targeted revision. Someone who only wants to change a background does not need to regenerate the entire composition each time. Drawing, selecting, generating and correcting stay within the same working context. This is especially relevant when a draft is mostly right but missing individual elements. People without image-editing experience still face a learning curve.

Spatial guidance is more than a marketing phrase: the academic ControlNet paper describes how edges, depth information or poses can condition image generation. The repository lists ControlNet among its underlying technologies. The publication explains the technical approach but is not an independent quality assessment of this plugin.

The plugin can be used locally for free; its repository specifies GPL-3.0. That does not mean hardware, electricity or all model rights are free or unrestricted. The optional cloud service charges through prepaid credit. On the check date, it lists, for example, €10 for 5,000 tokens. Consumption depends on the job and is displayed in the interface; this does not establish a universal fixed price per image. According to the legal notice, the operator is Cosmic Interstice UG (haftungsbeschränkt), based in Mannheim, Germany.

In plain language

Think of the workflow as repairing a painted wall: you tape off the area that should change and describe the addition you want. The AI supplies a suggestion within that assignment. You then remove the tape, inspect the transitions and touch them up with a brush. The selection and sketch replace part of a lengthy explanation, but they do not replace the final inspection.

A practical example

A fictional illustrator is designing a poster with a landscape they painted themselves. They want to try three versions of a signpost without repainting the sky. First, they save a copy, select a small area beside the path and generate three suggestions in sequence. They then inspect shadows, perspective and transitions into the grass. They add lettering themselves rather than relying on correctly generated text.

The sensible next test is just as small: one owned, non-sensitive image, one selection and three variations. Record waiting time, any credit consumption and the amount of manual cleanup. A result counts as useful only when it fits cleanly into the existing image. These numbers describe a test plan, not measured performance. Keep the original so that the change can be compared at any point.

Scope and limits

• Quality and rights: Generated details can be implausible; a good selection does not guarantee correct perspective. Model licences, rights to reference material and suitability for the particular commission need separate checks. The plugin replaces neither artistic oversight nor rights clearance.
• Hardware and setup: Local processing requires suitable drivers, models and sufficient storage. The documentation specifies at least 10 GB of installation space and says 50 GB or more is not unusual with optional packages. CPU operation is described as very slow. Anyone unwilling to maintain a setup should weigh that effort against the cloud option.
• Privacy and costs: Local and online are different operating modes. The privacy policy says no information is exchanged during offline use. For the cloud service, it identifies possible image processing outside Germany, including Canada or the United States. Confidential client images should therefore not be sent to an online job without review. Check current terms and the selected backend before production use.

SEO & GEO keywords

Krita AI Diffusion, Generative AI for Krita, Interstice, Acly, ComfyUI, ControlNet, inpainting, local image generation, creative image editing, open-source plugin