App support in ComfyUI

March 11, 2026By Alan Kent · AI agent architect; building Ordinary Animator

ComfyUI announced “App” support. It might not seem exciting to others, but I love it… because I need it and was doing a hacky approach before.

The ComfyUI idea is to allow people to share workflows more easily. The “inputs” to the workflow are clearly identified, meaning users don’t have to understand the insides of a workflow any more. So instead of users seeing a graph like the following:

Complex ComfyUI node graph workflow that users normally have to understand

They see a simplified UI with just a few inputs and the final output (the inputs are on the right with outputs on the left, which feels a bit backwards):

Simplified ComfyUI App mode UI exposing just a few inputs and the final output

The also have a new “hub” you can share workflows on. Makes sense. I mean my interest in this “App” feature is not because it makes a workflow work like a hosted App. My interest is because I want to chain workflows together.

More information on Apps can be found in their blog.

Workflow of Workflows?

My personal interest is in chaining of workflows, not “apps” as such. Why not just have a single big workflow merging different workflows together? Because I want a human to run a first workflow a few times, pick the best output, then proceed to the next step. I want a human in the loop.

For example (my domain is videos for storytelling):

A first workflow might be to create a background image from text.
A second workflow might then be to superimpose a character on the background.
A third workflow might be to re-light the image to be at night time.
And then finally that might feed into an image-to-video workflow.

The problem is (was!) that you had to open a workflow and work out which nodes inside the workflow to set to the new input value. Yes, you can make this easier by following conventions like putting the input nodes on the left or coloring them green, but inputs are actually at the property level, not the node level, so just identifying nodes is not enough - you want to identify specific inputs, and App mode lets you do that.

“App” mode is as simple as saying “these inputs of these nodes are workflow-level inputs”. You just click a checkbox! (You also identify the nodes which generate interesting outputs - the Save Image File etc nodes.) So you can say “yes, this seed property we should let the caller change”.

You can also add “labels” to App inputs to rename them. This is also important. If you have 2 input images, you have to give them different names to work out which is which. The node inputs might both be called “text” - labels let you call them “background image” and “character image” for an image edit workflow.

Apps vs Nested Nodes

Hey! Apps sound kind of like nested workflows! A nested workflow can have inputs and outputs. You can name them. Why didn’t they use the same UI?

(pause)

Heck if I know!

I will say however that nodes like Load Image today do not let you link to their input images, but you can mark them as App inputs. Since the goal is around making the UI simpler for users, there are probably technical details that made them come up with a completely new UI for what seems to be very similar to nested workflow nodes.

How might ComfyUI improve Apps further?

Crystal ball time! If I had to guess, the next step might be a node based interface chaining together “Apps”. Run one app, pick the preferred output, run the next app in a predefined chain. That is something I am exploring myself at the moment (I call them “strategies” to avoid “workflow of workflows”).

Fixed or dynamic workflow chaining?

I am also exploring the UI of understanding the ancestry of media assets. For example, the following graph was automatically computed by analyzing workflow runs. I now track the jobs media files were created by, and the inputs of jobs, allowing a dependency graph to be determined. Below is a simple tree, but its actually a DAG (Directed Acyclic Graph) as nodes can have multiple inputs (multiple parents). Still a work in progress (the media assets below are nonsense test images).

Automatically computed media asset ancestry dependency graph built from workflow runs

For me an open question is do I prescribe a series of workflows for the user, or propose a suggested list to help them solve a problem, or let them free form and do anything? For my use case of trying to help people tell stories, I don’t want to prescribe a flow, but I do want to guide non-expert users. So I was thinking of having a “strategy definition” that had a number of slots to be filled in - such as “background image”, “character 1”, “pose 1”, “lighting as text”, “first frame”, “last frame” (optional), etc. The user then fills out those slots incrementally until the reach the final output (e.g. a video clip for a shot in a scene of a screenplay). But still experimenting with what feels natural.