Logic of the text_encoder_training with Resume_Training

[Pasha49]

Hello! Please tell me how it will work?

If I want to train 10,000 steps, but with travel_text_encoder_for = 30%, and time limit for free training. How to do it?
In one training session, I can take 2500 steps.
30% of 10,000 = 3,000.

• 1st training(First_Training): 2,500 st and travel_text_encoder_for=100 (2,500 text_encoder)
• 2nd training(Resume_Training): 2500 st and travel_text_encoder_for=20 (500 text_encoder)
• 3rd training(Resume_Training): 2500 st and travel_text_encoder_for=0 (0 text_encoder)
• 4th training(Resume_Training): 2500 st and travel_text_encoder_for=0 (0 text_encoder)

or

• 1st training(First_Training): 2,500 st and travel_text_encoder_for=30 (750 text_encoder)
• 2nd training(Resume_Training): 2500 st and travel_text_encoder_for=30 (750 text_encoder)
• 3rd training(Resume_Training): 2500 st and travel_text_encoder_for=30 (750 text_encoder)
• 4th training(Resume_Training): 2500 st and travel_text_encoder_for=30 (750 text_encoder)

[TheLastBen]

Spreading the textenc finetuning through out the training steps should yield better results, go with the second one

[Pasha49]

Thank you! For two characters + the ability to use styles, how much percent will you advise?

[TheLastBen]

the characters must be from different genders, the text encoder must not go over 1000 steps, no matter the total training steps, then continue training the model without the text_enc until you get the desired result.

[Pasha49]

Thank you!

Total, for 2 characters of different genders. 15 photos on the character. Total photos 30.
30*200 = 6000 steps.
Maximum 1000 steps with Text Encoder.

• 1st training(First_Training): 2,500 st and travel_text_encoder_for=17 (425 text_encoder)
• 2nd training(Resume_Training): 2500 st and travel_text_encoder_for=17 (425 text_encoder)
• 3rd training(Resume_Training): 1000 st and travel_text_encoder_for=17 (170 text_encoder)
  Total 6000 steps, 1000 text encoder.

that's right?

[TheLastBen]

sounds right

[TheLastBen]

if you don't get the desired results, resume training for 200 steps text_encoder