Sharing the workflow I used to make comparison I posted on reddit. It's quite demanding to do it all in one go. It works on a 4090 but might work on other GPU, IDK. You NEED to "unload the models" after each generation or else, in my PC, it fills the VRAM and will get stuck.
Lumina is the most demanding. Even with the flash_attn installed. I get 2.5 it/s. It also don't accept a lot of resolutions.
I won't write right now how to install, look at each respective GitHub page. But if anyone needs, I can assist in the comments.
I updated the file with a more fixed version of the workflow. Now you can toggle what models you want to generate and the ones you don't.
Description
comparison Lumina/Pixart sigma/SD1.5-Ella/SDXL/SD3
FAQ
Comments (11)
SD1.5 Deep Shrink surprisingly beats everything. How is that possible?
Deep Shrink is really good to make 1024px images out of 1.5. So is Ella. Ella is fantastic... unfortunantly we won't get it for SDXL.
But to be honest, PixerlArt follows prompt really well in some examples. It's kind of all over the place in terms of who is the winner for me.
@diogod just seems to me we should be able to apply all this, ELLA, DeepShrink, PAG, SegMoe etc... to one thing and get super amazing results.
@diogod just waiting for a pony finetune of SD3 and ELLA / DeepShrink / PAG for SD3, then MoE SD3
I'm getting errors; I was told it's because this is out of date? Could you update the workflow nodes?
Could you please share a screenshot of the node with the error? Or the error itself? I would love to help (within of what I could do, I'm not a dev).
Also, would be awesome to include a MixOfExperts in this.
What do you mean by "unload the models"
On the ComfyUI Manager (on the side bar if you installed the manager) you have a button to "Unload Models".
this worflow loads too many models and ConfyUi will crash or overload the memory. On my comfy installation, the updated version, I can't even run it all at once anymore, I don't know why. I need to Mute half of the models. generate. Then (without unloading the models) unmute the rest and generate again. I used to be able to do it all in one go, and on the second run I needed to unload the models.