Krita AI Diffusion
Generative AI inside Krita: inpaint, outpaint, live-paint, and upscale without ever leaving your canvas.
About
Krita AI Diffusion is a plugin that brings generative image AI directly into Krita's painting workflow. Instead of exporting to a separate tool, you select a region and generate into it, erase an object with generative fill, expand the canvas with outpainting, or let the AI reinterpret your strokes in real time while you paint. It supports inpainting and outpainting, live painting, upscaling to 4K and 8K without running out of memory, ControlNet (scribble, line art, depth, pose, segmentation), IP-Adapter for style and face reference, and per-layer text regions.
The model story is refreshingly open: Flux 2, Z-Image, Stable Diffusion 1.5 and XL, and Illustrious all work, you can bring your own checkpoints and LoRAs, and everything runs locally on your hardware (a 6GB+ VRAM NVIDIA card is the realistic minimum) with an optional cloud path if your GPU is not up to it. At 10,700+ stars it is the de facto way Krita users do AI-assisted painting. The honest caveat is that local generation is still heavy: CPU-only or small-VRAM setups will be slow, and the install (Krita 5.2+, plugin import, model downloads) is a weekend-afternoon project, not a five-minute one.
What it can do
- Inpainting and outpainting: generative fill, expand, add or remove objects via selections
- Live Painting: AI reinterprets your canvas in real time as you draw
- Upscaling to 4K/8K and beyond without running out of memory
- Diffusion models: Flux 2, Z-Image, SD 1.5, SDXL, Illustrious plus custom checkpoints and LoRAs
- ControlNet: scribble, line art, Canny, pose, depth, normals, segmentation
- IP-Adapter: reference images, style/composition transfer, face swap
- Regions: per-layer text descriptions for different image areas
- Job queue, generation history, and style presets; fully local and open source (GPL-3.0)
Setup
- Install Krita 5.2.0 or newer from krita.org.
- Download the plugin from the GitHub releases page (latest zip).
- In Krita: Tools > Scripts > Import Python Plugin from File, select the zip, then enable it in Settings > Configure Krita > Python Plugin Manager.
- Restart Krita, open the AI Diffusion docker, and follow the setup wizard to install a backend and download models.
- For local generation use an NVIDIA GPU with 6GB+ VRAM (CUDA); AMD, Intel, and Apple Silicon are supported with caveats. Full guide: docs.interstice.cloud/installation.
Prerequisites
Krita 5.2+; for local generation, a GPU with at least 6GB VRAM (NVIDIA recommended); model downloads need disk space and a decent connection.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 10,724 GitHub stars (GitHub API), GPL-3.0, pushed Oct 3 2026; the README documents the inpaint/outpaint/live-paint/upscale feature set, the model list (Flux 2, Z-Image, SD 1.5/XL, Illustrious), and the hardware requirements; docs.interstice.cloud corroborates the install flow. Strengths: the standard AI plugin for Krita, genuinely local and open, strong feature depth. Docked for: (1) local generation demands a 6GB+ VRAM GPU in practice; (2) install is multi-step with large model downloads; (3) single-maintainer-led project. License: GPL-3.0.
★ 10,724 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: starScore from 10,724 stars on Acly/krita-ai-diffusion (log scale vs 298k peer max)
- · Maintenance 9: pushed 2026-10-03 (recency 10); steady release cadence on a mature plugin
- · Openness 9: GPL-3.0 (0.7*10) + product site and install docs verified at interstice.cloud (0.3*10); docked one point for copyleft vs permissive
- · Manual 8: de-facto Krita AI plugin corroborated by README feature set and docs; docked for single-maintainer lead and heavy local GPU requirements
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.