Introduction
Over the past few days, I've trained a handful of character LoRAs with Krea2, both real people and original virtual characters, fewer than ten in total. Here are some initial observations.
For training, I used Krea2 Raw as the base model with ostris ai-toolkit. All images were generated using Krea2 Turbo FP8. I haven't tested training on Turbo or with other training tools yet.
Key Takeaways
- Krea2 Raw is an excellent base model for LoRA training.
- LoRAs trained on Krea2 Raw perform very well on Krea2 Turbo.
- In terms of training stability, consistency, and generalization, Krea2 Raw is noticeably better than both Z Image Base and Z Image Turbo.
Performance, Consistency, and Generalization
I reused the same datasets and training settings that I previously used for ZIB and ZIT. No special adjustments were needed, and everything worked with Krea2 right away.
Training takes a little longer because the model is larger than ZIB and ZIT, but it's well worth the extra time. Unlike LoRAs trained on ZIB, which often lose noticeable likeness when used on ZIT, LoRAs trained on Krea2 Raw perform consistently well on Krea2 Turbo.
What impressed me the most was the consistency and generalization:
- Excellent facial consistency across different viewing angles, including ones that weren't present in the training dataset.
- Faces remain remarkably stable in full-body shots, without the obvious drop in likeness often seen with ZIB, ZIT, or FLUX.1.
- It doesn't unexpectedly shift the character toward East Asian features because of certain prompts, which sometimes happens with ZIT.
- Generalization is excellent. The character maintains a consistent identity across different outfits, makeup styles, and visual styles while blending naturally into each scene.
Current Issues
The only issue I've noticed so far is that characters sometimes appear slightly taller than expected, with body proportions occasionally looking a bit off. It's somewhat similar to the old Klein LoRA issue, although much less noticeable.
The Krea2 base model clearly retains recognizable features of some celebrities. When training LoRAs for those celebrities, the base model can still influence the results, even if their names never appear in the captions. I'll need more testing before drawing any firm conclusions.
One thing I can confirm is this: when using a well-trained celebrity LoRA at a strength of 1.0, including the person's name in the prompt tends to produce obvious signs of overfitting. For a properly trained LoRA, it's generally better to leave the name out of the prompt. On the other hand, if the LoRA is slightly undertrained, including the name can sometimes help reinforce the likeness.
Prompt Behavior
Krea2 handles prompts quite differently from ZIB and ZIT. To get the best results, prompts will likely need some model-specific tuning rather than simply reusing prompts designed for other models.
So Far
That's all for now. I'll share more observations as I continue testing.
I'll be releasing all of my Krea2 LoRAs on sololo.xyz over time, with selected ones also available on Civitai. Feel free to take a look if you're interested.
