With the Krea 2 contest, and thus training sale, ending in an hour I got what I wanted to convert (decided last minute to port a couple of unrequested styles that were "normal person" friendly because I thought they'd do OK with Krea 2's more realistic styles) and several requests done and have just one last Krea 2 model (not contest eligible) already trained ready for tomorrow. I think I’m going to be taking somewhat of a small break from model training for a bit after that final stretch used the last of my energy for now. I know I said I’d try a few things before the contest popped up (some Anima conversions, a few final Illustrious v2s, a couple of new things I wanted to try for Illustrious) but the contest had a time gate and needed first attention. Don't worry too much: I shouldn’t be gone longer than a week and don’t expect even that long, and I'll still be here genning images. Figured I'd do an AAR with my thoughts on Krea 2 and why I likely won't be coming back to it anytime soon (at least not unless someone pays my costs and accepts the results as they come, which I guess is possible with blue buzz early access) after I publish one last model (which isn't contest eligible) tomorrow.
On-site generator really unstable
I know this isn’t the model’s fault, but it really frustrates efforts to support Krea 2 when even just getting a successful gen is a roll of the dice. I regularly encountered estimates that kept getting pushed into the future till timeout, and a few times models that would just immediately error out for several hours after upload. It’s not even the time: I’d be fine if I could just submit and come back latter, but the error rate doesn’t allow that. I think part of this is loading the models into memory, but it just doesn’t make getting an initial gallery a nice experience. Even sample images in training wind up producing really weird artifacts like my actual image being a tiny thumbnail in a solid color field while those same prompts work fine after training.
Also it's more annoying to save my epoch pics now with the interface changes there (not a huge decrease, but it's annoying when that was already slower to do than I wanted). I am once again requesting staff add the ability to download all sample images at once as a ZIP.
It costs too much
Straight-forward. Just this last stretch cost my 6000 blue buzz between training and gallery gens, and that’s with a half off. I’ve got enough blue in reserve I could do it again, but I’ve got that reserve of blue by not spending ludicrous sums for unproven models, and would rather support something like Anima (which I couldn’t afford to port everything I made to even before this trek into Krea and needing to redo a bunch of stuff to fix Dim/Alpha settings).
I’m a tag user, it’s a “natural language” model
I’ve always felt my greatest strength as a model maker (and, to a lesser degree, image generator) has been having a solid understanding of Danbooru tagging and its idiosyncrasies before SD made it relevant. I’ve also long held the opinion that “natural language” too often devolves into tags but with extra steps because it’s impossible to truly know what kind of vocabulary the model responds to and its only parsing the prompt for certain bits. The big problem with it is while you can just look at the Danbooru tag list, with natural language there’s no dictionary you can look at: You’ve got no way to know, short of trial and error, if even specialist terms large segments of the literate native English public knows like “cuirass” or utterly normal phrases like “pointing at forehead with his thumb” are known to/understandable by the model at all.
The model is very weak at 2D without external resources
Now this one is absolutely a weakness of the model itself: Krea 2's main data and default is photorealistic, and I am not a 3DPD fan. It can do some ChatGPT "Anime" when specified but it's not good at that at all. While style LoRAs can bend Krea 2 into shape here, I’ve found even with that it its weakness at 2D means it struggles with the cartooning needed for “readable” expressions in 2D and only some basic ones can be conveyed easily. I doubt that's the only structural weakness it has for 2D either.
It’s ridiculously chaste
Also absolutely a fault of the model: Even for a model not trained on non-SFW content, Krea 2 is incredibly adverse to intimacy. Even something as modest as getting two figures to just hug (let alone kiss) proved absurdly difficult with my experience. Even just holding hands is a coinflip on if it will be out of frame. If it needs external resources to get even that I can’t really call it fully featured, and it really undermines one apparent advantage over older models of not fusing multiple figures, or seeming potential arbitrary posing natural language provides.
Seems a bit burnt in on certain poses?
Not sure on this one, but some poses seem way too repetitive across gens.
It does seem to grasp styles well I guess
Even with relatively limited data it makes a good facsimile of styles. Doesn't really overcome the other issues (especially emotion and posing) but I guess I can mention it as a positive.