working on image dataset and good description. also don't go over 6000 step for subject, and about 8000 for concept.

Good quality image is a default.

Using png with alpha channel and check "use alpha channel as loss weight" when train get difference.

I usually use 768x768 image for dataset. Seem get better results than 512x512.

Before training check model you are using, with simple prompt and no negative and really low cfg( 2, 3).

Check it with simple prompt, if result are something like your prompt, probally is good model for train, otherwise change it.

I use simple prompt like "man, green hair, blue eye, gold clothes". simple prompt, but with color mix to see what model better work on it. Some mode work good but only on certain cfg range, keep it low for me show how good model work in prompt.

Also made a merge of good model you found after check, is good idea to save a model just for use in training session. Remember to check every time you merge. Best choice for me is using model without any VAE baked in.

For the ratio of train, i use "0.02:200, 0.008:800, 0.002:2000, 1.06:3000, 0.0002:4500, 0.00008" for most train, some maybe need adjust, just check image result during train stage.

Another important thing in how you create embed. Initialization text and num.vector per token is important.

I use 8 vector for subject and 16 for concept. and same vector number for word Initialization text.

Btw this is how i work on Textual Inversion, and there are many other settings used by many other people's.

Just try and find what work for you, that require many time and fails ;)

@Potatovision i keep it disable in train concept, and at 1 for training subject, if subject is the same( like a person or a minifigure). for different subject( like different type of bottle) i keep disable.

@Potatovision happy training :)