When I was making my lora, I notice the characters eyes does not match the original style. This is a problem because it ruin consistency.

So I think the problem is the checkpoint used for training. And when I switch different checkpoint, the result is different, Turns out, that illustrious 0.1 is not the only good checkpoint for training lora like how people generally suggested.

I've got a hypothesis that "checkpoint trained" is more flexible than "checkpoint merge" when it comes to lora training. Because months earlier, I tried training a lora on wai illustrious 14.0 and the lora was not compatitable with other checkpoints.

So I think to improve my lora, I should use different training checkpoint. But it's much more complex than I thought. It has something to do with tags, the training checkpoint, and the generation checkpoint as well.

The best lora i've got is trained on Unholy Desire Mix - Sinister Aesthetic and then the images are generated with plant milk hemp ii.

Using any other checkpoint would lose some features.