HELLO! I am RAMTHRUST and I am pleased to present to you my daily driver workflow.
This workflow is something exceptional but VRAM heavy. It's so VRAM heavy that I think you need a MINIMUM of 16GB of VRAM.
It's slow, it's ugly, but godDAMN does it produce high quality images!
There are TONS of notes inside with download links to all of the detailers and other items you'll need to dance with the devil IN SDXL VERSION 1.01.
Basically, the process is as follows:
Get those missing resources, you should know how to do this in comfyui. There are download links in all of the notes for maximum convenience
native image gen - you know how to do this
preview chooser - choose the image(s) you want to upscale and hit proceed
If everything is in place, you should get a fantastic 2x upscaled image at the end.
There are multiple save points in this flow: The native generation, the latent upscale (on pass 2) and the detailer pass. I created so many save points because sometimes things go wrong and you may want to make comparisons without re-genning
IF YOU HAVE ANY QUESTIONS HOP INTO THE DISCORD GEN CHAT AND ASK FOR @RAMTHRUST and I will give you a hand.
I'm serious when I say that this upscaler is the bomb and I DON'T use anything else.
Please enjoy!
@RAMTHRUST
Special thanks to @CHESHIRE_OS because I'm a big fat front-end thief!
Description
Ladies and gents, I admit a failure on my part. I provided the an incorrect link to the face detailer that I use. It is supposed to be:
https://github.com/ultralytics/assets/releases/download/v8.2.0/yolov8m.pt
I am extremely sorry and will try to be better about this in the future.
FAQ
Comments (10)
I dont comment on things here often at all but i just wana say: WOW!
This is honestly an amazingly wonderful workflow. 10/10!
The results are just simply stunning!
Hey man, thanks for the kind words. I put a good amount of time and effort into this and it's nice to see someone else using it.
You have great taste!
Naturally just after release, I updated comfyui and now the preview chooser has lost some functionality in that you can no longer queue multiple jobs in a row. I am seeking a fix but probably won't be until the node is updated. The flow is still usable.
Have you considered using automatic CFG as part of this flow? I see it is set to 5.9 which usually works. But wonder if autoCFG would be even more versatile.
I haven't actually used autocfg get but it's on my testing list for sure
@RAMTHRUST I've generally had good luck with it. Mixed results with lightning/hyper-type models but normal and turbo models work well with it.
It works pretty well, better than Self-Attention and better than Perturbed Attention as well. I'll try to update soon as soon as the next best thing comes along
CORRECTION: I have run into MANY issues with autocfg after many hours of testing and nobody seems to be taking responsibility for the issue and I've been looking for months. Sorry man. But you were right, it is superior.
Issue is that it works sometimes, and others it fails out at the detailers with autocfg enabled. :\ total bummer.
But it's in the flow next to perturbed attention guidance, if you want to unbypass it, but bypass PAG if you do that.
This is excellent! Thank you.
Thanks man!