Krita AI Diffusion: custom ComfyUI workflow setup

This tutorial explains how Krita AI Diffusion sends image generation requests to a ComfyUI server running in the background. The presenter installs the server through the plugin, then uses the browser interface to edit workflows for local AI image generation. The listed demonstration hardware is an i5-14400F with 64GB RAM and an RTX 4060 Ti 16GB; the video does not establish minimum requirements.

The walkthrough addresses missing model files before exporting a workflow as JSON and importing it into Krita's graph mode. A Stable Diffusion 1.5 example runs after the presenter selects an available checkpoint. For Flux-2 Klein, the template expects different files from those downloaded by the plugin. The presenter selects an FP8 model and replaces the text encoder loader with a GGUF-compatible loader for the Q4 GGUF file.

Custom nodes connect the workflow to Krita. An output node returns generated images, a parameter node exposes the prompt field, and a canvas node supplies image dimensions and the seed. The presenter also explains why a fixed seed keeps producing the same image. The style node works with checkpoints that bundle the diffusion model, CLIP and VAE, but fails with the separate Flux files in this example. Sampling settings come from the plugin configuration rather than new input fields.