Trained on 37 photos on 七瀬カホ @capo130xx Instagram account. (If any legal issues are involved, please delete this model immediately.)

Trained on chilloutmix_NiPrunedFp16.

The effect of this embedding is not so good. The resulting picture doesn't quite look like her.

However, this embedding can indeed bring some changes to the face, please refer to the comparison picture I posted in the version "comparsion".

There will be no updates for this model in the future.

Some Thoughts

I will no longer be training Lora for the girls, only Texual Inversion for them.

Lora is a fine-tuning technique that affects the overall expressive ability of the model. And Texual Inversion is to explore the "prompt words" that allow it to draw "that girl" in the model's existing expressive ability.

So if you just want to make the face more like a girl, rather than completely replace with the girl, Texual Inversion is more suitable: it is lighter and easier to mix with other Texual Inversions.

As for face replacement, I think technology like ControlNet would be more suitable. Maybe face encoding can be the input codition for ControlNet?

I don't know much about it, so I won't discuss it too much.