I've long searched for a checkpoint based on IllustriousXL—a fine-tuned SDXL model that I could fully stand behind and use as my go-to. I’ve found a few that came close and have stuck with them for a while, but each one, as impressive as it was, came with a few glaring and frankly frustrating limitations.
After many failed attempts to create a cohesive SDXL-based checkpoint of my own, I decided to go all in. What followed was a long cycle of trial and error: dataset curation, Dreambooth training, merging experiments, and layering LoRAs and LoCONs until I arrived at this: v04, the current iteration of what I hope is a versatile, balanced, and usable base model.
Why v04? The journey began with me curating a mixed dataset of illustrations and photographs—personally handpicked based on what I look for: artists, poses, shapes, themes, and visual tone. I trained the first Dreambooth version on NoobAI, which I found to be a great base—it blends with different styles more fluidly, converges just as well as Illustrious, and avoids some of the strange quirks that tend to show up in Illustrious models. You probably know the ones I mean—hard to describe, but unmistakable once seen.
Here's the evolution:
v01 – the initial trained checkpoint on NoobAI using the custom dataset.
v02 – a merge of v01 with another model trained on an extended version of the dataset (with different repeat weights).
v03 – a combination of v02, yet another variant trained with tweaked parameters, plus three IllustriousXL-based finetunes.
v04 – a final merge between v03 and an experimental checkpoint made by merging NoobAI with several of my own LoRAs.
I’ve added a trigger word: "pkmerc" — it’s optional. It can improve result stability, but it’s not necessary. Many generations won’t differ much whether you use it or not.
One final note: the original base checkpoint that came out of the Dreambooth training (before v01) has the strongest personality and visual signature. It's bold and interesting, but unfortunately doesn’t handle LoRAs very well. I’ve decided to hold it back for now, but I may release it later as a standalone artistic model.