MixTrained-base1.0-test3需要触发词:@ogipote
krea2版本需要触发词:Ogipote style. 放在提示词首位或者尾部
MixTrained-base1.0-test3 needs a tiger word:@ogipote
krea2 version requires the trigger word: Ogipote style. Place it at the beginning or end of the prompt.
Description
FAQ
Comments (5)
What is the difference between MixTrained and NLTrained?
NLTrained was one of my early explorations into Anima's annotation method, using natural language annotations transcribed from booru tags. Based on current experience and mainstream recommendations, training solely with natural language annotations is not a good idea for preserving style during inference. Therefore, I am now focusing on optimizing and improving the training of MixTrained.
@ACatStick
thank you!<3
I’ll be happy to use it.
Based on my personal experience, the previous version was already quite good, avoiding some common issues often found in LORA (such as heavily imposing certain elements—about 70% of the LORA I've downloaded have this problem, or perhaps LORA itself tends to cause this tendency, though there are indeed excellent LORA that manage to avoid it). I also appreciate your continued research into better training methods.
I assumes that most LoRAs you mentioned were trained with too small dataset but wanting them to emphasize an abstract concept (like style or desired color palette), then too many dimensions were trained (Smaller dataset, smaller dim).