Using WaveSpeed models from Krita (AI Diffusion plugin)

This guide shows how to drive WaveSpeed-hosted models from inside Krita, using the Krita AI Diffusion plugin's Graph (custom workflow) workspace together with the nodes in this repository.

Nothing in the plugin has to be patched. The plugin can connect to any ComfyUI server ("Custom Server" mode), and it can run any ComfyUI graph that contains a Krita Output node. The nodes in this repo are ordinary ComfyUI nodes, so they work there as-is.

Two ready-made workflows are included:

File	What it does
examples/krita/wavespeed-image-edit.json	Sends the current Krita canvas plus a prompt to an image-edit model, returns the result
examples/krita/wavespeed-text-to-image.json	Text-to-image, fitted to the current canvas resolution

Both default to google/nano-banana-pro; swapping the model is a small edit, see Using a different model.

What this is not

This is not a one-click cloud setup. The Krita plugin talks to ComfyUI, so a ComfyUI instance has to exist somewhere you can reach it: on your own machine, on another machine on your LAN, or on a rented host. What you avoid is running a diffusion model locally — the WaveSpeed node only performs an HTTP call, so the ComfyUI machine needs no GPU for these particular workflows.

The plugin's own Online Service mode is a separate, unrelated service run by the plugin's author. This guide neither touches nor replaces it.

Prerequisites

1. Krita 5.2+ with the AI Diffusion plugin 1.26.0 or newer (custom graphs were introduced in 1.26).
2. A ComfyUI instance reachable over HTTP from the machine running Krita.
3. comfyui-tooling-nodes installed in that ComfyUI. This is the plugin's own node pack and it is mandatory — it provides Krita Canvas, Krita Output and Parameter, which are how images and settings cross between Krita and the graph.
4. This repository installed in the same ComfyUI (see the README), with your WaveSpeed API key configured under Settings → WaveSpeed in the ComfyUI web UI (or in config.json). The key lives on the ComfyUI machine, never in Krita.

About the plugin's connection checks

When connecting to a custom server, the plugin verifies that four node packs are installed (ControlNet Preprocessors, IP-Adapter, External Tooling Nodes, Inpaint Nodes) and that at least one usable diffusion checkpoint is present. On a ComfyUI that only has the WaveSpeed nodes, that check fails and the connection is refused.

If your ComfyUI is a full local install used with the plugin already, there is nothing to do. If you are setting one up purely as a relay for WaveSpeed, you can relax the check: close Krita, open settings.json in the plugin's user data folder — typically %APPDATA%\krita\ai_diffusion\ (Windows), ~/.local/share/krita/ai_diffusion/ (Linux), ~/Library/Application Support/krita/ai_diffusion/ (macOS) — and set:

"check_server_resources": false

One caveat we could not work around: the plugin still fails with "No diffusion model checkpoints found" if the server reports an empty checkpoint list. ComfyUI needs at least one file visible in models/checkpoints for the connection to complete. These workflows never load it.

Note that the plugin's other workspaces (Generate, Upscale, Live, …) keep using local models. Only the Graph workspace runs the workflows below.

1. Point Krita at your ComfyUI

Enable the docker (Settings ▸ Dockers ▸ AI Image Generation), open the plugin's Connection settings, choose Custom Server, enter the server URL (for example http://127.0.0.1:8188) and connect.

If your ComfyUI sits behind an auth proxy

The plugin supports a bearer token, but there is no widget for it — it is written into settings.json (same file as above):

"server_authorization": "your-token-here"

The plugin sends it as Authorization: Bearer your-token-here on both the HTTP requests and the websocket connection, which is what most reverse proxies expect. Restart Krita after editing the file.

Do not put your WaveSpeed API key here. This token authenticates Krita to your ComfyUI; the WaveSpeed key is configured inside ComfyUI.

2. Import a workflow

1. Download wavespeed-image-edit.json or wavespeed-text-to-image.json.
2. In Krita, switch to the Graph workspace in the plugin docker.
3. Click Import File and select the JSON.

The workflow is copied into the workflows folder of the plugin's user data directory, so it remains available in later sessions.

The files are in ComfyUI's API format, on purpose. The WaveSpeed node builds its widgets dynamically in the browser, and a UI-format export does not survive the plugin's UI→API conversion (that conversion maps widget values positionally against the node's declared inputs, and this node declares none). If you author your own WaveSpeed workflow in the ComfyUI web UI, use Export (API), not plain Export.

3. Generate

After import you get a small parameter panel, generated from the Parameter nodes in the graph:

• 1. Prompt — the edit instruction, or the description for text-to-image.
• 2. Aspect ratio — one of 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9. Pick the one closest to your canvas.
• 3. Resolution — 1k, 2k or 4k.

Press Generate. Under the hood:

1. Krita sends the visible canvas to ComfyUI (edit workflow only).
2. The WaveSpeed node uploads that image, calls the model endpoint and waits for the result URL. Expect a few seconds to a minute — it is a remote API call, and the plugin's progress bar cannot show fine-grained steps for it.
3. The result is downloaded, scaled to the canvas size and handed back to Krita, where it appears as a result you can apply as a new layer.

Because the models' aspect-ratio choices are coarse, the returned image is resized to the exact canvas dimensions by an ImageScale node at the end of the graph. Choosing an aspect ratio close to your canvas avoids visible squashing.

Cost and rate

Every generation is a billed API call against your WaveSpeed key. Keep the Graph workspace in regular mode rather than Live — Live re-runs the graph on canvas changes, which for a hosted API means a request (and a charge) each time.

Using a different model

The model choice lives in three inputs of the WaveSpeedAIPredictor node:

• model_id — the API path, e.g. /api/v3/google/nano-banana-pro/edit
• request_json — default values for that model's parameters
• param_map — parameter types, including which ones are arrays (this is what folds image_0, image_1, … into an images array)

The reliable way to switch models is to build the graph once in the ComfyUI web UI:

1. Open ComfyUI and drag in wavespeed-image-edit.json.
2. On the WaveSpeed node, pick another model. Its widgets update automatically.
3. Reconnect the Krita Canvas image output and the prompt Parameter to the new model's inputs if the names changed.
4. Export (API), then import the file in Krita.

Editing model_id / request_json / param_map by hand also works, as long as the parameter names match the model's schema.

Structure of the included graphs

Krita Canvas ──image──► WaveSpeedAIPredictor ──output_url──► WaveSpeedAI Preview
     │                        ▲                                      │ image
     │        prompt / aspect ratio / resolution                      ▼
     │                   Parameter nodes                         ImageScale ──► Krita Output
     └──── width / height ─────────────────────────────────────────► ▲

• Krita Canvas supplies the canvas image plus its width and height.
• Parameter nodes become the input fields in Krita.
• WaveSpeedAI Preview turns the returned URL back into an image tensor.
• ImageScale fits the result to the canvas.
• Krita Output is what makes the graph importable at all — the plugin requires at least one.

Known limitations

• Selections / inpainting are not wired up. These graphs replace the whole canvas. A Krita Selection node can be added for masked editing, but the models used here take a full image plus a prompt rather than a mask.
• Animation mode is not supported by these graphs.
• No sharing with the plugin's styles. The model runs on WaveSpeed, so the plugin's Style settings, samplers and local LoRAs have no effect here.
• Progress is coarse — the job simply shows as running until the API call returns.

Feedback

Problems with these nodes or workflows: please open an issue in this repository. Problems with Krita or the plugin itself belong in the plugin's repository. This integration is maintained by WaveSpeed and is not affiliated with, or endorsed by, the plugin's authors.