Comfy-Code

A simple framework for generating and running comfy graphs from typescript.

Fully typed nodes. Write entire comfy graphs without ever having to open the ComfyUI Web Interface.

Getting started

This project requires node version 23 or newer!

npm i comfy-code

Make sure your ComfyUI Server instance is running.
Then generate ComfyUI Typescript classes (you will need to run this command every time you install new Nodes in ComfyUI)

npx comfy-code import nodes

This should have created an imports folder in your current directory, which contains classes for each and every node in Comfy.

By default comfy-code will expect your server to run on http://127.0.0.1:8188.
If you use different settings, check npx comfy-code import nodes --help for options.

Writing your first graph

(See the examples folder in this repository for complete examples. The t2i example is explained below)
Use

const activeGroup = ComfyNode.newActiveGroup();

to get an array which will automatically store all subsequently created ComfyNodes.
You can omit this if you want to keep track of all your nodes yourself.

Start by creating a node which loads a checkpoint, like:

const loadCheckpoint = new CheckpointLoaderSimple({ ckpt_name:'checkpoint-name' });

Your IDE should give you the ability to auto import the CheckpointLoaderSimple node from your imports folder. Otherwise you must do so manually like

import { CheckpointLoaderSimple } from "/imports/loaders/CheckpointLoaderSimple";

If everything works correctly, ckpt_name should have intellisense which reflects your currently installed models.
When you create a Node like that, the arguments represent the incoming connections into that node. They can either be primitive values (such as a string or a number) or references to outputs of other nodes.
Let's create our clip text encoder nodes next so you'll see what I mean by references:

const textEncodePositive = new CLIPTextEncode({ text: "positive prompt", clip: loadCheckpoint.outputs.CLIP });
const textEncodeNegative = new CLIPTextEncode({ text: "negative prompt", clip: loadCheckpoint.outputs.CLIP });

Here we used primitive string values for the prompt's text, but we connected the clip input to the clip output of the load checkpoint node we created earlier.
Another more dynamic way to create connections would be

loadCheckpoint.outputs.clip.connect(textEncodePositive.inputs.clip);
// or
textEncodePositive.inputs.clip.connect(loadCheckpoint.outputs.clip);

The rest is just more of the same. We create EmptyLatentImage, KSampler, VAEDecode and SaveImage nodes and hook their sockets up to one another. Because this is javascript, we can use if clauses or for loops to build a more dynamic graph much more quickly than in ComfyUI's Web-UI.

And at the end of this all, our activeGroup array will have automatically stored all Nodes, and we can test it in ComfyUI. For this we need to create a ComfyInterface instance

const comfy = new ComfyInterface('http://127.0.0.1:8188');

This ComfyInterface exposes all API routes which ComfyUI makes available to us.
In this case we want to run our graph, so we call

const promptResult = await comfy.executePrompt(activeGroup, "print");

The print option lets us see progress updates in the terminal.
If you did not use activeGroup, replace it with an array containing at least all the output nodes. Comfy-Code will infer all dependency nodes by itself.

Now all that's left is to run the script, which I use ts-node for.
If you do not have ts-node installed globally, install it locally

npm i -D ts-node

Then run the script:

npx ts-node ./src/test.ts

And that's it. Your generated graph is being processed in Comfy.

Importing an existing workflow

You can turn a json workflow file into a typescript script by using the comfy-code import workflow script. Both api exported workflow files and regular workflow files are supported, but the api version is preferred.
Example:

npx comfy-code import workflow -i ~/Downloads/Unsaved\ Workflow.json -f -o ./test/workflows/workflow.ts

Use the f flag to generate a script which will execute the graph as a prompt when run. Omit the f flag to just generate the graph.
Use --help for more options.

Let's quickly run the workflow to see if everything worked:

Then run the workflow script:

npx ts-node ./workflows/workflow.ts 

The default -f setup should now print progress updates to the console.

Scope/Future of the project

This is a side project. Pull requests that improve existing features will be merged. Bugs will be fixed, feature requests will likely be ignored.

I view the project in its current form as feature complete.