lora training
#6
by luxdelux7 - opened
Thanks for all the work
i was thinking of training a few more 3d style loras on top of the model with something like
https://github.com/gazingstars123/Anima-Standalone-Trainer
since that gets decent results fast
could you perhaps give a few tips for both style and character training aka since the model expects the specific format
how would one format the fields in each txt when it comes to style?
so for description i assume a proper full text description? or also tags somewhere as well?
cheers
Thanks! Note the README covers the format, not an official LoRA recipe — but the arch is identical to Anima-Base v1.0, so any Anima trainer drops in. Just match the caption format.
Each .txt = the structured caption, same field order:
{
"year": 2025,
"preference_level": "best",
"artist": [...],
"copyright": [...],
"character": [...],
"image_description": "...",
"extra_tags": [...]
}
- Style: style trigger →
artist. That's the field's purpose. - Character: character tag →
character; series →copyright. - Description: yes, a full proper natural-language description. No tags in it — tags go in
extra_tags. The model was post-trained for natural language at the 512-token scale, so prose is the default. - Keep
year/preference_levelconsistent across the dataset, and caption the same way you'd prompt at inference. - No quality words or negative prompt needed (negative = empty).