Pixart Sigma - Like ELLA, but for Low VRAM SDXL+ Resolutions in ComfyUI!

Pixart Sigma was released recently, and while the main repo takes a little tweaking to run nicely on 24GB VRAM, ComfyUI comes to the rescue making it easy to run in just 6 GB!
Like with ELLA, T5 encodings replace CLIP leading to increased prompt adherence. Inthis video I show you how to get going for FREE in ComfyUI, and compare SDXL vs Pixart Sigma.
All this without one of those SD3 “no commercial use” licenses!
Want to support the channel? Get workflows and more!
/ nerdyrodent
Links
github.com/PixArt-alpha/PixAr...
github.com/city96/ComfyUI_Ext...
huggingface.co/spaces/PixArt-...
github.com/PixArt-alpha/PixAr...
huggingface.co/PixArt-alpha/P...
== More Stable Diffusion Stuff! ==
* Installing Anaconda for MS Windows Beginners - • Anaconda - Python Inst...
* Installing ComfyUI - • How to Install ComfyUI...
* ComfyUI Workflow Creation Essentials For Beginners - • ComfyUI Workflow Creat...
* Faster Stable Diffusions with the LCM LoRA - • LCM LoRA = Speedy Stab...
* Make an Animated, Talking Avatar - • Create your own animat...
* One Image Gets You a Consistent Character in ANY pose - • Reposer = Consistent S...

[@MrSporf]

Wow. I'm impressed. The prompt understanding is something else - quite amazing!

[@NerdyRodent]

Yup, it’s pretty good!

[@valerymoyseenko]

Thanks buddy! Especially caught your words that PixArt generates more different faces than SD. These identical smiling cute ladies with almond-shaped features are already starting to get annoying, and it has not even been a year. I have installed PixArt at my place, I'm studying it now, but the first results are impressive.

[@steffen_r]

Thx for sharing. The demo looks promising and the model has a good understanding of prompts.

[@NerdyRodent]

Yup! It’s great for prompt understanding and so ideal for generating an initial image 👍🏽

[@user-eq1rf2hq8s]

Ah, so that's why I couldn't load the model in comfy! Thanks - you cleared up where I was going wrong with the custom node install.

[@NerdyRodent]

Glad I could help!

[@Armaan27012012]

do you have a workflow which has pixart and controlnet and ipadapter?

[@Digital_Toolbox]

Thanks for the great content! What I wasn't sure about was if I already have comfyui Install up and running, can I just skip the preparation step and move directly to step two? Thanks!

[@NerdyRodent]

Yup, the video assumes you’ve got ComfyUI installed already 👍🏽

[@harshitpruthi4022]

i am trying to do some experiments , will it work with either ipadapter or faceid , have anyone tried it

[@bebeno1]

I wish we could convert any fine-tune sdxl model to T5 encoder like juggernaut that would be amazing

[@WatchNoah]

what linux distro are you using / can recommend?

[@NerdyRodent]

Ubuntu LTS 😀

[@WatchNoah]

@@NerdyRodent nice, thanks for the quick reply

[@l.x6]

lol the end

[@WhatsThisStickyStuff]

This uses SDXL components but isnt actually sdxl right? Like we cant use sdxl control nets or anything

[@NerdyRodent]

Yes, you need to use Pixart control nets

[@fredmcveigh9877]

Hey Nerdy. I respect your opinion and you haven't said anything about SD3. Please let us know what you think of it and the fact it doesn't seem to be open source anymore.

[@NerdyRodent]

my focus is on things with open licensing, so with the SD stuff all being for research only, non-commercial use now it’s less of a focus. Great if you’re ok with paying the 20 dollars a month though!

[@fredmcveigh9877]

@@NerdyRodent It's just sad that they seem to have changed their philosophy on people being able to use their product for free. Blender for example still continues to be free to the user. If the usage was reasonably priced then it wouldn't be such a problem but when you experiment and fiddle around with prompts like i and many others do it will be very expensive.

[@LouisGedo]

👋

[@NerdyRodent]

👋

[@maslaxali8826]

6min left

[@JustFor-dq5wc]

This new model is great. Thanks Rodent! .git directory in pixart_sigma_sdxlvae_T5_diffusers have extra 20 GB. Can be deleted.

[@Hooooodad]

Hoping this comes to forge

[@CapsAdmin]

I'd guess they are samey on sdxl because you are using a community model. I think it would be fair if you used the base model, at least in addition.

[@NerdyRodent]

There is, of course, the actual prompt understanding component to consider as well

[@ngones7280]

I feel super sad when I see something very cool and is Cuda dependent, fuck AMD

[@NerdyRodent]

Linux + CUDA is where it's at for AI at the moment :/

[@ngones7280]

@@NerdyRodent ye.. :/

[@San4itos]

I use ROCm on Linux and I'm Ok with it.

[@ngones7280]

@@San4itos RX580 here so I don't know how to use Rocm properly and think HIP isn't available I even try to use Zluda in a new version and message related to HIP shows up