Trained on Krea 2 Raw using a curated dataset of 22 images. This LoKr adds only marginal value (slightly more vibrant colors and less clutter). Krea 2 Turbo already generates convincing Kinkade-style images out of the box. Here's a side-by-side comparison:

Recommended settings:

• 8 steps

• cfg 1.0

• sampler: er_sde

• scheduler: simple

Prompting:

Trigger word is not required but use Thomas Kinkade style or similar wording for best results. For typical Kinkade scenery (cabin, village, snow, etc), short prompts in the form of comma-separated phrases will suffice. But for non-typical use, long prompts produce much better results.

Known issue:

Outputs may contain signatures when using long, natural-language prompts mentioning art style and medium. This comes from the base model's pretraining, not from the LoKr or dataset cleanup failures.

I personally use Mage Flow Edit to remove the signatures. Check out the second showcase image for an example workflow.