Finally got my Krea2 LoRA trained, now it can actually nail those detailed art styles, not just waifus anymore.

I’ve trained a few different LoRAs myself, and after testing them, this is the first model that actually manages to reproduce the details of a refined fine-art style. Honestly, it kinda blew my mind. A lot of older models would just default to anime waifus or fake influencer photos when you try to push a style, so the range of what you can do with them is pretty narrow.

Krea2, on the other hand, has a noticeably wider scope—it can handle those more pure art styles that emphasize brushwork and texture. The images I posted here are all single-pass generations—no upscaling, no inpainting, no touch-up passes, just straight out of the box. That’s a big deal for teaching demos, because a one-shot image makes it way easier to explain the logic.

Sure, the consistency for single-image generation still needs more testing, but the starting point here is already way higher than the previous batch. Anyone else who’s trained style LoRAs, feel free to compare parameters.

“It’s not just about the 2D waifu” — that line hits way too many people.

Damn, the quality straight out of a single generation is seriously impressive.

It’s a big step forward if it can handle a pure art style.

No upscaling, no inpainting, makes it even more impressive.

Anyone got some training parameters I can reference?

For teaching demos, straight-up generated images are definitely easier to explain.

Same here, especially curious about how many images are in the dataset. For style stuff, I can never get the quantity right.

same here waiting on params, never dared crank the learning rate on style LoRA