Training for model based on SD1.5 got bad result

[steelywing]

I have tried training on several SD1.5 base model (Ex: ChillOutMix, NED, ...) for person, but all get bad result. If I use the same dataset to train on SD1.5, the result is good, even use the textual inversion on other SD1.5 base model the gen image also good.

I am using SD WebUI to train.

Image train on SD 1.5, then generate ChillOutMix:

Image train on ChillOutMix, then generate ChillOutMix:

1 of training dataset:

[rinongal]

Are the finetuned models you are training on also realistic models? Or are they more restricted (e.g. to artistic styles)?

[steelywing]

Are the finetuned models you are training on also realistic models? Or are they more restricted (e.g. to artistic styles)?

The model I training on is also photo realistic model, but A model itself will produce a face with A character, B model will produce B face, as far as I understand (they should have finetuned the face), textual inversion should get the face weights better on their own model, isn't it?

I can train on SD1.5 with no issue, I just want to understand why it get worse result, I am thinking if I am not setting config properly, thanks for your reply~

(I edit the first post for clearer descript the train and generate model that I use)

[steelywing]

I found if I change the vectors to >= 16, the result a lot better, I see some YouTube guide says vector 2~4 for character training is enough, may be he only trained on SD basic model, thanks you