Custom workflows

Every canvas action runs a ComfyUI workflow. Most features can run your own workflow instead of the built-in one, for example a different image model or edit model. ComfyArt finds the nodes to fill by their title: you title the nodes that receive the image, prompt or size, export the workflow in API format and upload it. You can do this yourself or let an agent do it.

Let an agent do it

Ask the ComfyArt agent, or any agent connected over MCP, something like “connect this workflow to quick edit” and give it the API-format file. It reads the rules with the custom_workflow tool, titles the nodes, checks the workflow and installs it. The rules are also a built-in skill, Connect a custom workflow. For an agent that is not connected to ComfyArt, open Custom workflows in the project menu and copy the rules with the button next to the agent hint.

Do it by hand

1. Title the nodes

In ComfyUI right-click a node → Title and add a marker:

Marker	Node	ComfyArt fills
`@image`	Load Image	the selected image
`@image2`	Load Image	a second image (style reference)
`@mask`	Load Image (as Mask) or Load Image	the area to change, white = change
`@last`	Load Image	the last frame of a video
`@prompt`	a node with the text as a widget: CLIP Text Encode, TextEncodeQwenImageEdit, String (Multiline)…	the prompt
`@size`	a node with `width` and `height`, e.g. Empty Latent Image	width and height (and `batch_size` when there is no `@count`)
`@width`, `@height`	Int	width and height, when they are separate
`@count`	Int, or a node with `batch_size` / `amount`	how many results
`@seconds`	Int / Float, or a node with `seconds` / `duration`	video length in seconds
`@save`	Save Image, Save Video, Preview Image…	the result. Needed only when the workflow has several end nodes

Titles are case-insensitive and other words are ignored, so Positive @prompt works. A node that receives several values takes several markers: @prompt @size. Each marker goes on one node only. When the text of a CLIP Text Encode comes from another node, put @prompt on that node. When a video workflow counts frames instead of seconds, add an Int node titled @seconds and turn it into frames with a math node. ComfyArt randomizes every seed and noise_seed on each run and finds models in subfolders, so these need no marker. ComfyArt’s own workflows in custom_nodes/ComfyArt/workflows/ are titled the same way, so one of them is a good starting point. Other @ names in them (@sampler, @decode…) are internal and ignored for custom workflows. The built-in expand workflow pads the image inside the graph, so it is not a template for expand.

2. Export in API format

In ComfyUI choose Workflow → Export (API). A normally saved workflow (with a nodes list) is rejected.

3. Upload and assign

Open the project menu (top-left icon) → Custom workflows. Features are grouped into image generation, image editing, image tools and video. Click the upload button next to a feature: ComfyArt checks the markers, says what is missing, and uses the workflow for that feature once it passes. Each feature keeps its files in its own folder, comfyart/workflows/<feature>/ in your ComfyUI user folder (edit, upscale, video…), and its menu lists only those. Pick Built in to switch back. ComfyArt never deletes your workflows; to remove one, delete its file in that folder; a feature whose file is gone goes back to built in.

Markers per feature

Feature	Required	Optional	Notes
Text to image	`@prompt`, `@size` (or `@width` + `@height`)	`@count`	Appears as an extra model in the generator. Reference and base images are not passed.
Quick edit / edit text	`@image`, `@prompt`		Also used by compose and mockup blending.
Style transfer	`@image`, `@prompt`	`@image2`	Without a reference, `@image2` gets the same image.
Inpaint / edit region	`@image`, `@prompt`		ComfyArt crops the region for your workflow and pastes the result back. The result node needs an `images` input.
Expand	`@image`, `@mask`, `@prompt`		ComfyArt pads the image first; your workflow gets the padded image and the mask of the new area.
Upscale	`@image`		ComfyArt resizes the result to the chosen factor. The result node needs an `images` input.
Remove background	`@image`
Eraser	`@image`, `@mask`		Also fills the background behind edited elements and split layers.
Multi-angle	`@image`, `@prompt`		The prompt is a camera description such as `<qma> front-right quarter view eye-level shot medium shot`.
Split layers	`@image`	`@prompt`, `@count`	`@count` is the number of layers.
Video generation	`@prompt`, `@size`	`@image`, `@last`, `@seconds`	Used by the first/last frame mode. Reference, motion control, lip sync and storyboard keep the built-in model.
Animate image	`@image`, `@prompt`	`@size`, `@seconds`, `@last`	With loop on, `@last` gets the source image again.

If ComfyArt has an image to pass (a first frame, a last frame, a style reference) and the workflow has no marker for it, the run fails with a message instead of ignoring the image. Storyboard, video extend, creative upscale, move object and video upscale always use the built-in workflows.

Example: use another edit model

Open your edit workflow in ComfyUI.
Title the Load Image node @image and the node with the edit instruction @prompt.
Export in API format, upload it, and pick it for Quick edit / edit text.

Get started

Create

Assets and Agent

Let an agent do it

Do it by hand

1. Title the nodes

2. Export in API format

3. Upload and assign

Markers per feature

Example: use another edit model

​Let an agent do it

​Do it by hand

​1. Title the nodes

​2. Export in API format

​3. Upload and assign

​Markers per feature

​Example: use another edit model

Let an agent do it

Do it by hand

1. Title the nodes

2. Export in API format

3. Upload and assign

Markers per feature

Example: use another edit model