Liselotte Cretia from Seirei Gensouki aka Spirit Chronicles. Liselotte finally appeared in mature form in anime S2, so there is a lot of new material for training now.
Versions before v3-screencaps-illu:
Liselotte Cretia from Seirei Gensouki trained on mostly official light novel art.
Trained it from scratch using new settings and manually processed some images in the dataset. Especially hair should be much more consistent now compared to previous versions.
Description
This version is trained without pruning captions, so it needs a lot of trigger words to create Liselotte consistently. It tends to work less consistent without hair bow in particular.
Trained with NAI (12 epochs x 18 images x 15 steps), examples generated with AOM3a1.
All examples used openpose contronets and upscaled with multidiffusion plugin.
FAQ
Comments (3)
I'm very pleased that someone did this Lora.
I also have several questions and would be grateful if you could answer them.
From what I see, you're using a powerful graphic card Can you specify what card you are using for generating images?
Why are you using multiple epochs for a training process?
Isn't it distorting the result?
Did you try only one epoch and compare it with this one?
I'm using RTX 3090 with 24G of VRAM indeed. Have been using 3080 with 10G until recently and I could train loras with it too, couldn't upscale images as much though.
About epochs, they are just a way to count steps. You have number of images times steps per concept times number of epochs You can also save snapshots after a certain number of epochs (which don't affect training), I usually do so as often the final snapshot tends to be overbaked and you can test one with lesser number of steps if you save it.
Well, this one doesn't look overbaked after additional 24 epochs (compared to LiselotteCretia4), will probably upload one more model soon. Although to make a real difference, I think I'd need to work more on image set.
@faux hmm, according to the Automatic1111 guidelines, an epoch isn't steps but a set of actions(works) that need to be performed.
Anyways, if it works for you, then that should be fine. I was just curious that maybe you know something besides the information stated on Automatic1111's guideline.
Thanks for answering my questions. It gives me a bit of food for thought and a reason to think about upgrading my 3070 8GB to a better GPU.
I will be waiting for your new Loras and updates for this one.
Kind regards,
KUZY2X