Time for some finger lickin' BB-Q pitlickin' fun at the ol' glory hole!
Krea2 v1: Created using a spruced up version the Flux version's image dataset and retained previous captioning.
On a side note, it seems that Krea2 learns complex concepts pretty easily, at least faster and without a LR tinkering compared to Flux or Z-Image. Pretty much out of the box settings and it started picking this up very fast during the training epochs.
As with the current models, the magic will be your prompting and not just relying on the trigger word to do your heavy lifting. Just a few word changes or slightly different phrasing can make or break a successful image generation. Krea really seems to pay close attention to the prompt so if ti's doing exactly what you want, just revisit the prompt and be more precise (and/or try a different seed) - it'll listen!
FLUX version: I couldn't get FLUX to get this concept quite just right with out a lot of very extensive prompting and even then it still required extensive img2img regeneration edits. So, I created this as a helper based off my SDXL dataset for the same concept. I cleaned up and upscaled the previous, salad tossed in some new ones and what came out of the FLUX oven wasn't too bad so I thought I'd share.
While this version isn't perfect (for example, it tends to make everyone in the initial image generation w/ their tongue out, not to mention how most FLUX base models don't handle adult situations well out of the box), it still does a decent job even with other LoRAs stacked in. I've observed in my testing that it can generate pretty good results in a one-shot, about half the time, but that'll depend on your prompting, checkpoint and other generation parameters. If the output has artifacts or deformations then usually it takes just one or two img2img corrections to get it right. ...
TL;DR: As a 1.0 version so don't expect a one-money-shot miracle every time, but just img2img correct with one or two generations to a final version.
SDXL version:
keyword: pitlicker
Not exactly satisfied with the model because about half of the time it produces unusable results so probably will make a 2.0 version .. this version will work for the most part if you mess with the strength and depending on the complexity of the prompting or other LoRAs involved.
Made this one a LoHA and I am experimenting with different types so if I figure out a better way to get this model consistent while playing well with other LoRAs with flexibility for complex prompting I'll update it. Including training images too in case you have a better approach and I'm still learning..