Optimal parameters (unet, text encoder, concept steps)

[specblades]

Before I have had too many unsuccessful attempts (4 so far), I want to ask - how can I calculate in advance what parameters to set?

I would be insanely happy if you could share your experiences.

For example.
I train for subjects.
Different representations of the same class of objects, 40 images, each named with 4 tags, 200 concept imgs
tags: men/women, armor, leather/chain/heavy/futuristic, helmet/helmetless

As I understand it (for the next attempt, prev was totally garbage), the approximate values should be: 4000 unet, 1200-1600 text encoder, 600-700 encoder concept steps.

Correct me please.

[TheLastBen]

don't use tags for the images, use only one single identifier for all the instance images, and for the concept images, you don't need to rename them or crop them, just upload them.

for 40 images and 200 concept images, start with 3500 UNet, 600 text_enc, and 800 concept_text_enc

[specblades]

This time I trained a model with these parameters: 32 images, 4 tokens of 8 images, 3000 steps, 480 text encoder, 580 concept encoder. Tell me if I set the parameters wrong.
Next time I will try 20 images, 2 tokens of 10. But in this case, I seem to lose the ability to point generate the right type of armor. I dunno

[specblades]

Oh, I think I'm beginning to understand, variety is achieved through training the concept encoder, and the instance images should not be a lot

[TheLastBen]

exactly, you let the concept images help with the diversity, and the instance images with the visual style.
the concept images can be as small as 256x256, you can upload up to 1000, make sure they fall in the same category as your subject for example "warrior", "knight" ... etc

[SAC020]

don't use tags for the images

why not use tags / caption the images?

[TheLastBen]

it will cause overfitting