Introduction
I've been training Lora styles (and some recent models) in the SDXL family, including Pony and Illustrious, for about two years now. However, with time and experience, I decided to create my own trigger tag that adds the Lora to the model without losing style details. Depending on the model, some Loras work without problems without the trigger tag due to the strength (weight) used in the Lora.
However, all this changes in the Anima model, which is a completely different family from SDXL, using tags and natural language.
Tests
I noticed that in this latest release (Chelodoy Style Lora), when trying to train with the trigger tag "by chelodoy," the Anima model didn't apply the style correctly and basically left the default style. When replacing it with the trigger tag "@chelodoy," the trained style was automatically strongly linked to the Lora, but the training was compromised because the tag (by chelodoy) was present both as the main tag and in the characters (this is a standard practice of mine).
Upon checking the model's hugging space, I noticed that Anima behaves differently when the "@" symbol is added during LoRa training (especially for style) if there are more than 50 images trained on the base model. To better explain this situation, I will give examples:
1) In the LoRas I recently created, "Shredded Wheat Style" and "TOTO (Inaxtoto / bebebebe) Style LoRA", you can see that I used the trigger tags "by shreddedwheat" and "by toto inaxtoto bebebebe"
2) If you test and add the "@" to the "by" in the base Anima model, you will see that there are few changes, very small ones, which I deduce means that there are not enough images used in the training of the base model for these artists, therefore the standard trigger tag "@" will not have any influence on the LoRas.
3) After 10 training sessions with various configurations in the Chelodoy LoRa, I noticed this change, and looking at the number of images trained in the base model, there are more than 50 images, which may indeed have some influence when using the "@" tag.
Conclusion
Due to these changes, I decided to standardize the inclusion of "@" in style tags according to the template author. I could have left it only for artists who received more images in the training on the base Anima template, but it's annoying to keep guessing who will receive the "by" or "@" tag.
For those who want to follow and learn more about this subject, I'll leave the links below. It's quite interesting because I didn't see many of these topics on Illustrious/Pony, for example.
List of artist styles used in the Anima model:
https://thetacursed.github.io/Anima-Style-Explorer/
About the use of tags / list of artists / tag weight:
https://huggingface.co/circlestone-labs/Anima/discussions/6
https://huggingface.co/circlestone-labs/Anima/discussions/14
https://huggingface.co/circlestone-labs/Anima/discussions/112
https://huggingface.co/circlestone-labs/Anima/discussions/73
https://huggingface.co/circlestone-labs/Anima/discussions/164
Note: I may ultimately be wrong about this, but since the model is very recent, there may be better and more in-depth explanations on this topic. However, in my tests I was able to perceive this difference; if anyone else has noticed this or knows of any other explanation, I would be happy to read your comment.