Smolder | Character LoRA (Artist : Kanel) | Pony & 1.5 Reviews

Hi, really wanted to leave a commentary about LoRA. First of all, i've checked LoRA metadata and noticed, that you had 49 pictures, but only 20 on them was tagged. That's not the way you should do LoRAs i believe. Or you just got 20 1girl-tagged Smolder images in folder somehow.

If you're not tagging images in your dataset, training process starts to train all object on your image in pair with training subject, like it was a part of them. For example, you have a lot of untagged outdoor images with clouds. If you prompting your characters, there is a big chance of clouds appearing from nowhere.

I think you've forgot to tag images when you've released enhanced version of LoRA.

Just an advice. If you're making LoRA with training resolution of 1024x1024, you really want to scale down your images to 1024x* or *x1024, by the smallest size. So, you image resolution should look like 1024x1768, for example. This really improves LoRA quality and makes buckets less funky. The same for 512x512, 768x768 etc.

That's all, i guess.

This is the only Smolder LoRA for now, so i'll wait for improvements, thanks, by the way.