I downloaded your dataset to try and figure out why this LoRA failed so spectacularly. And at first it looked ok, until i checked out your "prompts". And those prompt are a recepee for disaster, and will never produce a good LoRA. If you are planning on training another LoRA, dont use keywords as prompt for your images.
The text encoder for Krea2 is a language model, so it is not trained on keyword. That will only make things worse, as the only relation the text encoder has with keywords are anime images.
If the prompt you have on a image dont produce a similar image if you just generate a image with it, the prompt will do more harm then good. The prompt should produce a image that is as similar as possible to the dataset image it is linked to.

I like the LoRA. At around 0.49 it does exactly what it’s supposed to do.

For me anything up to ~0.60 is usually the sweet spot — if you push it beyond that, you’ll likely fry the latent space with silhouettes from the training images. Depending on the dataset, that can look like the model just came out of a blender.

I also saw another comment mentioning problems. In many cases that kind of “everything turns into mush” result simply comes from running the LoRA at strength 0.8–1.0 or even higher. Krea2 LoRAs don’t behave well at extreme strengths; they collapse into the training silhouettes very quickly.

On top of that, I think there’s a misunderstanding about how Krea2’s text encoder works. Krea2 is not a BLIP/T5-style language model and it doesn’t require natural-language sentences. It uses a tag‑cluster embedding system, which is why tags are actually the recommended format for Krea2 LoRA training — that’s also what Civitai advises.

The text encoder is frozen during LoRA training, so it doesn’t learn grammar or sentence structure. It only produces embeddings that act as labels for the diffusion model. Whether those labels come from tags or from a structured sentence doesn’t matter — what matters is consistency.

BLIP-style chaotic sentences don’t work well, but they will also work using structured prompts like “photo of a woman doing X at Y wearing Z” behave like ordered tag containers. They produce stable embeddings and Krea2 can cluster them reliably. That’s why tag-based datasets (or syntactically stable sentence prompts) work perfectly fine.

So the idea that “keywords will never produce a good LoRA” is not accurate for Krea2. Tags are not a recipe for disaster — they’re the intended method. Most issues people see come from running LoRAs at extreme strengths, not from using tags. So stay creative — and don’t be afraid to use tags when training Krea2 LoRAs. They’re exactly what the system is designed for.