That's odd, there shouldn't be anything special about LUSTIFY! that would make it work when other SDXL models fail.
The only special thing it has I can think of is this blurb:
"This isn't a lightning, but DMD2 model. For those who don't know, it's purpose is the same as Lightning/Hyper, but with a different look."
Based on other models it taking ages to load (10+ minutes), and then stalling out shorty afterwards, my first thought would be a memory issue. Given that it is the same size as all other SDXL models, I don’t see why it would be different, though. That said, the one OnlyForNSFW one worked, it did when I had very few other programs asking for RAM at the same time. (Only Firefox with a few tabs and my notes app.)
IDK how to do that in ComfyUI, but maybe someone else here knows how. If Comfy is too confusing, you might want to switch to A1111.
Awesome to see that you found something that works for you.
Just to be clear, I’m using ComfyUI because the installer worked, not because I’m particularly attached to it. I had issues installing both InvokeAI and Forge, though I’m pretty sure what went wrong with Forge and could probably get that going.
ComfyUI works okay once it is set up. However, everything feels very needlessly inconvenient. All these dozens upon dozens (hundreds, really) of
tiny nodes for every little thing. Why the heck do you need half a dozen nodes for the KSampler? Why are does (almost) every single need to be loaded extra instead of just one big node for the basic things? It feels incredibly bloated.
Of course there
are more conventient nodes. After all, every user can program their own nodes. This just confuses the situation. Take the "Load LoRa" node.
The basic node lets you load one (1) LoRa. Want two LoRas? make more nodes. Suddenly, you have 10+ nodes just for the LoRas.
So someone made a LoRa stacker that lets you load three (3) LoRas at once. Why three? Who knows. You still need to fully unload a Lora to turn it off, though, so have fun clicking around a lot if you want to test things!
Then there’s an expanding node that can fit as many LoRas as you want, but it only lets you adjust the strength, not the clip strength. (Shouldn’t matter, though, since most don’t affect clip anyway.)
Then there’s a LoRa node stacker that has an on-and-off button. Convenient. But it doesn’t work alone and instead requires you to load
another node as well to get it to work. (I think so, at least.)
And so on…
For basically every single thing you might want to do, there are different impementations, often with different settings and different requirements. And given that the whole ComfyUI ecosystem is based around sharing workflows with other people, it’s one huge mess where every one does the same things differently.
Long story short, I
think have a correct implementation for the Face/Lips/Hands/etc Detailer now, but it also seems wrong because it didn’t work. Or to be more precise: It took
48 minutes to attempt to fix the hands in an image that took ~4 minutes to generate in the first place, and it just made them worse instead. I have about a dozen decent pictures with just one awful detail*, and my motivation for figuring out how to fix it is very low. I’ll probably try to install Forge again tomorrow, or maybe go for A1111 and use the baked-in Detailer that one is supposed to have.
It’s very annoying because I feel like have everything else essential figured out by now. I’m even pretty sure I could do the ControlNet stuff if I really wanted to for specific poses. The only other thing I failed at is making the suspended spitroast LoRa work, and I think that might just be an issue of it being too cartoonish for photorealism.
*Well really, I also have a couple of nightmare fuel images I accidentally made that I want to make presentable. Probably something to post myself once I have P4, though. As a hint: If you start your prompt with "evidence photo of", the scenes sometimes look like places the police might be called to…
