Calling ComfyUI Workflow Nerds~

Coming up this week I'm working on porting Comfyui inside of my training ecosystem. I have never really been previously a ComfyUI user. I've just downloaded like 100 workflows, but you know for a fact that I'm not going to probably understand most of these.

I could in theory ask my peers, but what works for an onsite generator for Arc En Ciel may not work for a single user UI.

Why?

"WHY" ARE YOU DOING THIS? you ask? - Because i'm actually as you know, done with Gradio for the most part. I'm not willing to move into FULL COMFYUI node use, but I'm willing to use ComfyUI as a "BACK END" - as a tool to make my work much easier. Also because I know so many are done with Forge forks not working with previous extension architecture.

Plus honestly, originally my idea was to fork FORGE and rip it to shreds, but the community around forge gets really VALIDLY opinionated. I don't blame them, they want COMFYUI compatibility but with that familiar UI. But as NEO changes all of the structure which is REALLY COOL and NEAT (but severely frustrating to me in short lol) - it becomes hard because trying to update existing plugins means "AINT NOBODY GOT TIME FOR THAT"

What?

"WHAT ARE YOU ASKING FOR" - I'm asking for people that make workflows to come up with stuff that you'd want me to test making a UI around. I have a template and I have an Idea of what i'm going to be doing, but being that Next/Node is REALLY fluid (reactJS) - I'm looking for both advanced and and simple.

But i'm also looking for people who are interested in conversations around how these work, how I can optimize this to work for anyone who's got 8gb vram all the way up to people who are insane enough to use multi GPUs.

Er ok? But SD -SCRIPTS?

Yea it's also a training ecosystem, I don't mind having MULTIPLE tools at people's disposals YET making it simple enough to just MAKE a choice to either FLAG install it one way or even turn something on or off in the settings.

My background is mostly in WEB and Graphic design, so you'll note that my UI is a bit all over the place but we're getting there. I'm not a professional developer, but Next/React/Node is both web and back end linguistics in coding, and so it's MUCH FASTER for me to design around Claude and Qwen helping me. Unlike python, which is not something BAD it's just i'm not good at python.

So my hope is this is a broad spectrum HUB/ECOSYSTEM that's useable in ways for anyone in a way but specifically people like me to a point.

So I know in my LAST article nobody understood based on no screenshots what this thing looks like right now - it WILL be changing a bit we just moved to Next 16- i'm learning MORE AND MORE each week about what we can and can't do.

SCREENSHOTS!

(warning some are a little out of date)

WHERE?

Where do I find this monster of a toy?

https://github.com/Ktiseos-Nyx/Ktiseos-Nyx-Trainer/

Yea here, now you have the screenshots!

What TYPE BTW?

Again looking for workflows that YOU would like to see in, and to be quite fair I know i downloaded a TON but I'm curious to see what people are willing to suggest -- or even create specifically.

Your workflow will be connected to a UI, credited in the readme as well as the attributions page.

HOW CAN I HELP?

Well if you don't usually make workflows but you use them, you can help out by documenting how a lot of these work. You can develop matching SHADCN/RADIX components to any custom nodes you'd like to be able to eventually use as plugins to the system.

Plugins MAY NOT WORK out of the box the first month or so - but the point is to eventually slot in custom nodes or workflows like you would either A1111 or ComfyUI.

If you're into developing plugins, and want to sketch out and plan some neat stuff feel free to join up the discord here: https://discord.gg/HhBSM9gBY

WHAT DOES IT RUN ON

The system itself barely takes more than 200mb of ram even while training, being that my own windows ENV only has 8gb of system ram. (yea i'm still on that imac, just bootcamped into Windows.) So the whole point of using Next was to limit the amount of ram leaks, which sure ComfyUI may still need more to run - but less overhead means MORE wiggle room.

I can't train on 8gb of VRAM because I don't have enough system ram overhead to handle the mess, plus let's be real - i'm not on a standard PC.

It is compatible with REMOTE, Windows, Linux (and likely *Nix/BSD) and MacOS. MacOS is literally just going to be determined by: ONLY SILICON machines, because I highly doubt ZLUDA/ROCM works on older macs. I'm not saying it doesn't, but if an LLM still thinks Microsoft is the reason for the season, and forgets to tell me it's not just because i have a non SSD -- then it's up to us to test these things.

I've tested it on Runpod with both an older A series card and 4090. I'm trying to remove a lot of the bias around requirements FOR a 4090 - because I'm aware people have way more patience then I do. I also know for a fact it runs on VastAI: https://cloud.vast.ai/?ref_id=70354&creator_id=70354&name=Ktiseos-Nyx-NextJS-Trainer

There's still a lot of incremental annoyances, that in theory should be "SMOOTHED" before I go into breaking everything. But i'd rather take a break from that and break it all in half, make it explode and re-formulate it in a way that works.

So right now, WINDOWS it should really work - but because my DEV machine works, and Zero who consistently tests it for me uses FORCE install, we don't have many other windows testers - "MY MACHINE WORKS" theory doesn't help because I have everything neatly installed. (*Insanely messy and FFXIV is slow because I forget it hates less than 1tb of HDD)

So the true answer is : A potato.

A potato with at least 8gb of system ram, and at least 8gb of Vram. I'm sure you could try less, because everyone says you can but it doesn't mean you should.