optimal parameters for the subjects

[svumme]

these days I tried to train the model with 55 images of a subject but I can't get a good result. The best test is this but not satisfactory:
I set UNet_Training_Step=6000
and Text_Encoder_Training_Steps=350
Text_Encoder_Concept_Training_Steps:0
and style_training not checked.
I also tried 12,000 steps and 20,000 but both went into overfitting.
Could you tell me how to calculate the various parameters based on a given number of images? I would be grateful to you.

[TheLastBen]

reduce the instance images to 15-20, too many instance images confuse the model, and set the training steps to 3000-4000

[TheLastBen]

So you're probably loading a different model that the one you trained, and also the V2 give awful results with faces, use the v1.5

[svumme]

probably it is as you say, there were over 100 models :) i was reading on the checkpoint field of the webgui but maybe something was not working, because now it works fine. for 1.5 do I have to upload a yalm file? and which one is correct?
thanks

[TheLastBen]

I don't think you need to add a yaml for the v1.5

[svumme]

Thanks for all the replies, and sorry for the doubt I had. Now it works perfectly, the 1.5 is very good. I have a feeling the yalm works better. But it's just a feeling I have to do more tests.

[jujpenabe]

Did you tried training with v2.1?