Hi! Pretty much the same—I haven't quite gotten the hang of training ANIMA LoRAs yet, but I just gave this one a try.
Description
Comments (13)
The @ at the start is used for artist names (it's a way to let the model tell style tags apart from everything else). It's not a strict rule, but I imagine the same dataset without the @ might result in an even better lora
I'll give it a try
Maybe I should mention that in my guide for anima training settings?
You can tag away style to some degree with "@", e.g. @3d.
@RisingV Absolutely! You can add it to your guide. I think using @ for a character was a mistake, and using the character's original name as the trigger is risky too — if ANIMA already knows the character, the base model's priors can interfere with or override what the LoRA learns. The safest approach: a unique alternative name for the character, and @ reserved for artist/style names.
Your lora seems to work fine with omitting @ and just using "jackie lynn thomas". Personally I haven't had any problems with interference for any of my character models, I used "nami" for Nami, "may" for May, "nico robin" for Nico Robin, "rui (pokemon)" for Rui (though official tags for Nami and May are "nami (one piece)" and "may (pokemon)". But yeah unique names will work fine too.
@RisingV Thanks for testing the model! Good to know the name without @ works fine — I'll keep that in mind for future LoRAs. Appreciate the insight!
No problem. I guess your LoRA also works without the trigger and just using character traits, since you did not use the trigger in the showcase images?
@RisingV Yeah, I tried it both ways. It works without the trigger, but with it the features stay more locked in — without it sometimes traits drift or get a little inconsistent. So the trigger is less of a hard requirement and more of a "tighten things up" kind of thing.
@RisingV Also, the trigger word isn't showing up simply due to how I wired my nodes in ComfyUI, so Civitai isn't picking them up. Wasn't intentional, just how my setup ended up working out.
@Blitzkralle1505 I see...
@Blitzkralle1505 Why not use both the character's original name AND a unique name/trigger for maximum flexibility?
@nogo I think it depends on whether ANIMA already knows the character. If it does, using the original name can reinforce what the model already has, while a unique name gives you a separate "version" without messing with the base knowledge. If ANIMA doesn't know the character at all, it doesn't really matter... Either way the LoRA learns from scratch. So having both isn't really about "maximum flexibility" for the LoRA itself, more about whether you want to work with or around the base model's existing knowledge.
@Abel1502 Sorry for starting a whole thread under your comment, didn't expect that to happen hahaha