Trained on the style of ForestTheRotten along with some synthetic data to attempt to create the ultimate style for the gooberest of goobers with goober faces. I trained it locally using https://github.com/ThetaCursed/Anima-TrainFlow and it worked brilliantly.
-
Trigger: ForestZeRotten
Recommended Weight: 0.8
-
Sometimes I feel
Like I’m a hot goblin
Little bit perfect
Little bit problem
Description
Comments (6)
a small typ for training new styles into anima
add an @ before the trigger word so that the usual naming convention takes place. makes the lora more dynamic
What do you mean by more dynamic? What would that actually do? I haven't seen any other Anima loras doing that yet, so it doesn't seem to be a convention for trigger words at least. I do know @ is used for artist name tags in Anima's base training, but I haven't seen it used for lora trigger words yet. It might even be confusing for people to see @ in prompts thinking it was an artist name tag when it's actually a specific trigger word for a Lora. Anyway, I don't mind doing it going forward if there is actually some reason for it, but I just don't see it.
@Razira Hmm, I did see it in several style LoRAs. My assumption is that using @ may help because Anima already associates that token pattern with artist/style identifiers in the base model training. Since the LoRA is trained on top of those existing representations, it should make trigger align more naturally with learned style concepts and reduce collisions with normal words.
For example, if someone makes a literal Rotten forest scene, your trigger world would clash
@Valadion Oh I see where the confusion is. I didn't use the actual artist name as a trigger word and just picked a similar unique tag that isn't in the Anima dataset. If you wanted to double up with Anima's artist training and this Lora you would use my trigger word ForestZeRotten and the Danbooru set artist name with @ which I believe would be @forest box. I specifically avoid training on the artist name so that I don't clash with the model's base artist training. You can use them both together for a heightened effect usually, but it can cause unintended effects. As an example, Vanilla Cookies has a fairly desaturated look which I try to offset in training. Using the @ for the artist would give you a desaturated look. Yuni does a lot of extreme contortion which I exclude from training. Using @ for Yuni will probably lead to distorted anatomy. Ideally the style lora provides a style without leaning on an existing tag.
@Razira Yeah I've seen those loras with the @ of the style trained on the base model but I don't get why would that be any better than have them separated, if wanted to merge both style trainings from base and the lora style then do so, but if not, then the trigger word being the same one used for the style trigger word would be useless for that purpose, I think even Anima fails to completely show the style desired, a lora works better for that, also Anima base probably has lots of styles but merging all of an artist style image into the style, even skeetches, unfinished drawings etc, while a lora can have only the finished drawings and stuff and probably a lora works way better to show the proper style than the style trained on the base model, can be dynamic by just using both at the same time, but using the same trigger word to me is pointless and not useful at all if wanted to use them separately, so that's why I think it's better to have a separate trigger word or tag.
@Suomsoh I've played around for some of the built in artists and it does well for some, but falls flat for most. It's mostly a matter of having enough specific training. A good lora will always be better. Fortunately, Anima is even better than Illustrious at picking up on styles during lora training. It's quite a nice model!