Version 2.0 is trained with different dataset with less muscle size and definition. It's also trained with different settings to make it more flexible. 1.0 had some bias to change art style

@_1_ This is embedding, not LoRA. Embedding creates an unique token word that AI associates with certain attributes/vectors during training. Training is very random and is a bit of luck. Sometimes AI learns features you don't intend it to. In my experience you can increase the chance of AI making the dataset images style into a vector by increasing number of tokens per vectors amount. There's a script that lets you see the loss and magnitute of vectors of embedding. If magnitute is around 0.2 then embedding starts to ignore other words in prompt and becomes 'inflexible'. Training for long time, using high learning rate and setting a low vector per token number makes magnitute increase very fast. So need to play with this numbers and try to hit around 0.19 vector magnitute at end of training. If magnitute is too low like 0.13 then it is most of the time under trained and can't generate features consistently. Embedding training is difficult but I think it's better than LoRA in terms of flexibility.

@_1_ For training characters I use 25 images of character naked with simple background. I use around 15 closeup face photo and 10 portrait style with torso. No need to show legs. It makes AI create vectors for body features as well. In caption I include the background color and naked so it doesn't associate those with embedding keyword. In Initialization text I give AI a simple starting point, briefly describing the character in prompt style. For Keira Strong I wrote 'big burly woman with red hair' as initialization text. I use 5 to 8 vector per token depending on complexity of character. If your character has very unique features, AI has hard time learning your character, so need to lower vector amount to prevent it from being distracted by a lot of character features. For generic man, woman characters 8 works good. Learning rate 0.001:30, 0.0005. Batch size 1, Gradient Acc. Steps 1. Max steps 500. I save an image and copy every 10 steps. With this numbers my training hits around 0.19 magnitute at 500 steps. But keep in mind it doesn't mean 500th step is the best. When training is over I compare the early copies with XYZ script to find the embedding that has learned the features I want best.

@AITSCH inspect_embedding_training