The style of Rtil.
Trigger word is @rtil or @xtilxtil depending on the version. The latter uses a different tag than what the model knew previously, so the knowledge of the style should be purely based on the dataset. I'm not sure which is better. The images were tagged with the year and decade (10s or 20s). You might want to put sketch, simple background and/or patreon username, watermark in the negatives. The dataset is the same as my Illustrious version, except the changed trigger word. The model is a Unet-only DoRA as usual, I didn't have the time to experiment with the config yet.
Turns out I have plenty of time for training if I run it overnight. And bear the fan noise. But yeah. Anima Base 1 actually knows the style pretty well already, but I found it starts to struggle at higher resolutions, so I decided to update the LoRA. Btw, to be clear, all of my LoRAs are in line for updates. Once I'm done with the toolkit for dataset prep and checkpoint post-processing, I should be able to put out a LoRA every two days. Just you wait...
Next is probably gonna be pettankon. I feel like it mixes really well with this style, and I'm curious to see if the effect is any different on Anima.
19.09.2026 - Added regular LoRA version in addition to the DoRA.
If you like the style, please support the original author.
Description
I realized DoRAs had downsides I hadn't previously considered, so now I'm remaking LoRA versions of my DoRA models. In short:
DoRA works better alone and on the same model it was trained for
LoRA works better in combination with other adapters or on checkpoint finetunes
LoRA also appears to work better on Anima 3.8B
FAQ
Comments (3)
What is Dora
One of many evolutions of the idea of a LoRA. From a user's perspective it's practically the same thing. The difference is in the math of how it is applied. Check this out if you actually want to know the details. To put it very simply, it can learn more complex stuff in a smaller file, but it slowly grows worse when the model it is applied to changes (for example when you use a different finetune, or combine it with other loras).
If you've heard of LoCon, LoHa, LoKr, LyCORIS, etc., that's another one of these