Overview
A LORA for Animagine XL V3.1. Trained using OneTrainer.
Usage
Trigger keywords: 1girl, sakimiya iruka, dolphin wave
Recommended keywords: green eyes, red hair, one side up, hair ornament, green scrunchie
Recommended strength: 0.5 ~ 0.6
Description
FAQ
Comments (12)
2.7 GBs for a lora?
I tested various dimensions with OneTrainer for various characters and so far 256 felt the best. It's probably not optimal, but it's been consistently working.
Back when training SD1.5, I also felt like high dimension was the best despite what people say. So, for now, I seem to be stuck with using this methodology until I stumble upon a better way to do it.
@CulturedDiffusion To be honest, when approaching lossy method (which ML training did) a higher rank/dim yield higher quality although the increase is puny small (that many people said doesn't worth the extra cost, I disagreed). However, pardon the tangent, considering SV3D or Tooncrafter uses 20-48 GB VRAM, a low rank of under 4 GB is well, seems "low" rank. For another case, there is also tonade who add <1000 imgs into his t-pony-nai and saved as huge checkpoint, but personally I think checkpoints are trained if the data ranges 100K-1M images, less than 1K is more into LoRa realm. What I want to say is, if <1000 imgs become 6-8 GB is justified to avoid being lossy, then LoRa of 2.7 GB seems acceptable.
@trytensor I think whether it's worth it or not also depends on the use case.
A lot of LORAs focus on capturing the main features of the characters across many different art styles. This works fine if you want to draw the character in the base model's art style/s.
However, I personally like to train LORAs that recreate not only the character but also their official art style. This has been difficult to achieve consistently without raising the rank.
@CulturedDiffusion just a little advice , you can get a consisten way of recreating a character artstyle if you only use screenshot from their anime, it doesn't matter the size all that much, the dataset has to be pruned of Fanarts, it has worked for me but sometimes getting the right amounts of screenshots it's a pain in the ass Ofc i'm working with pony so i don't know if animagine or base XL work differently
@Chichiue_Pendragon Yes, this is how I've been working with LORAs since the beginning, even back in SD1.5. With that said, some characters have very little official art (such as new gacha characters), so this method has unfortunate limitations.
Great! The artwork is very close to the original style. Could you create Suminoe Shion, who was voted the most popular?
I don't think I'm going to train other Dolphin Wave characters anytime soon since there are a lot of characters from other franchises that I'm interested in training.
@CulturedDiffusion I understand because there are so many characters in this world. Thanks to the internet, there are more characters to create than one could complete in a lifetime. I look forward to seeing more of your works.
@masakilin Hi, what model did you use to get closer to the original style?
@KPNK Hi!usually, just use the one recommended by the creator.
@KPNK The LORA was trained on Animagine XL 3.1 so it should give the best results: https://civitai.com/models/260267/animagine-xl-v31