Prompting:
Generally, start with "Oil Painting of ..." with the LoRA at 1.0 strength and you'll get the desired style for most situations.
However, some subjects, and longer prompts, cause the style to break down. Figures may be depicted in 19th century clothing when not desired. Or faces and other subjects will be overly refined in a digital-illustration style.
To counter this, reiterate painting terms. For instance, end the prompt with "Classical painting techniques in oil on canvas depicting [general subject genre], loose gestural paint style." Simplify or expand on this to try and push Krea2 to stay on target.
Alternatively, raise the strength of the model, though this can get ugly too often, so try re-prompting first.
If you want NSFW output in a painterly style, pair with a lightweight LoRA that uncensors the Krea2 model.
About the style:
My interest in training style models started with an embedding I made for SD2.1 - called Classipeint (clever naming is not a strength of mine) which got much better with a LoRA for SDXL, ClassipeintXL.
The intent was to bring a classical painting style something like a combination of Courbet, Sargent, Turner, and similar painters near the end of French Academy period of the 19th century. Pre-impressionism, but leaning away from the high realism of someone like Bouguereau.
I remain very proud of the SDXL style. When we got more capable models in Flux and Z-image, I hoped to bring the same aesthetic to them.
I consider my attempt with Flux to be a failure. That model can't get past its own style, flattening textures to random noise or photographic imagery.
Z-image was much better than Flux! And with the greater prompt adherence, was a big success to me. But still, Z-Image didn't quite learn the ideas of brush strokes and surface texture as thoroughly as I hoped for such a sophisticated model, and it worked against a lot of images by introducing both more sharp detail, complexity of detail, and vibrancy of color (characteristics of digital illustration more than traditional oil painting) than I wanted.
Krea2 initially seemed like it would fight me as well, but showed enough promise in my first attempts (using the original SDXL dataset) that I kept at it. Eventually, I developed a new, higher resolution dataset, and through a series of merges of various training runs, finally hit on a balance that, I think, is the most satisfying yet. This has that energy and atmosphere of the SDXL model (in fact, even more convincing rendering of brush strokes and paint marks) but the greater prompt adherence of Z-image.
Limitations:
Though probably the best realization of the painting style I've been chasing for a these few years of model training, it does still fail in a number of scenarios:
prompts for a generic girl tend to output a 19th-century girl
long complex prompts tend to lose the aesthetic, becoming more photographic or digital-illustration
references to magic often result in more digital illustration styles
I'm sure it breaks down in other scenarios - these are just the ones I've seen the most so far.
refer to my initial statements, above, on prompting to mitigate these failings. So far, I've only had a small handful of prompts where I genuinely cannot render the scene I want with the painting aesthetic this LoRA is intended to represent.