Krea2 training two characters at once, faces keep blurring together

Q&A / Help / 求助答疑
1

Need some technical help here—I don’t really get the underlying stuff, but I’m stuck on this. I’m trying to train two characters at the same time in Krea2, gave each their own keywords, used the full model (not the turbo one), but in the preview the two faces slowly blend together. Same exact issue I had with Z-Image before.

Then I tried training separate LoRAs for each and using them together during generation—still no luck, every face ends up as a mix of both.

Back in SDXL, I could train three characters in one LoRA and they’d come out perfectly distinct. Why doesn’t it work in Krea2? Is this a common flaw with these newer models, where faces just naturally merge? Or am I doing something wrong? Any vets here can point me in the right direction?

2

It seems like merging faces for multiple characters is pretty common with the new models.

3

Training them separately and then using them together still causes bleeding, I’ve run into that too.

4

Honestly, the tricks you learned from SDXL don’t always carry over to new models.

5

Try using regional prompting to separate the different parts when generating.

6

The keyword distinction isn’t strong enough, so they tend to blend together. Try cranking it up and see what happens.

7

Z-Image does the same thing, probably a problem with that type of architecture.