Got sick and tired of trying to figure out how to get a so-called "Victorian Anime" aesthetic, which I'm enthralled with, so I made a LoRA. Used Counterfeit v 2.5 as the base model. It's my first time making a LoRA, so I have no clue about a lot of things regarding them, but I used 20 images to create this one... probably should add more images later, since I noticed that a great majority of the ones I used were close-up portraits, and not full-body stuff.
Anyway, try it out and see what works!
Monet
Update v 2.0: Saw that Counterfeit ver. 3.0 dropped, so I attempted to update VAA to see if it works well with the new model. Doesn't seem to be the case, as it turns out deep-fried and with a persistent green artifact [Update: I wasn't using a VAE; best to use one to avoid this issue]. But, VAA v 2.0 seems to work nicely with Counterfeit ver. 2.5. Perhaps I should wait for a pruned version of Counterfeit 3.0, and see if VAA will meld with it then.
I typically throw weights to the extreme (2.0 for certain prompts, and 1.0 for VAA) to see what happens. Oddly enough a weight of 1 on VAA doesn't seem to produce bad results for me on Counterfeit ver. 2.5. But I would recommend no more than 0.8 if you're working on a different model.
Update v 3.0: Trained v3 using the dataset used to train VAXL, with some slight modifications to the prompts here and there. I upped the Network Dimension and Rank by a factor of 2, and as this dataset is the largest of all the SD1.5 VAA LoRA, it's a bit hefty. Now it includes some trigger words, namely "Victorian", "Renaissance", and "Monet".
Description
FAQ
Comments (4)
A decent idea but this looks very overfit. If you want, contact me on twitter and we could work together to create a larger dataset..
https://twitter.com/Elf_Anon
I used to make a lot of images like the ones you used for training and can help you make more. It's kinda sad how modern models don't have the same artist weights so I think I understand where you're coming from.
Yeah, I'm very new to all of this, so I'm sure I did something wrong. By overfit, I presume you mean how they seem similar to each other? Tried to rectify it in version 2, by adding more images for training... kinda got a mixed bag of results... probably made things worse by doing so. I'm also afraid to say, I don't have a twitter account, but I'd definitely appreciate advice on how to make this LoRA better. Would definitely appreciate having better images to work with for training too XD
@Monet_Einsley Overfit means it's too similar to the source material. Like if you train an anime model and typing "1girl" gives you a character from that anime, instead of a character in the style of that anime. In this case I could clearly see the source of training instead of just seeing a new and unique Victorian babe 😘
It's not really a big deal, but I think more generalized models that deviate from the source are better overall.
@Dweeb Thanks for the insight! Will work on it. Gonna try and generate lots of good source images to train v 3.0