Bad training results with multiple subjects

[bach777]

I used 4 different subjects, with unique names, and high quality/cured 30 photos each using the recommended settings, however the results with multiple samplers, cgf scale and steps, do not resemble each other at all and do not preserve the facial features and body dimensions corresponding to the input. It is still better to train individually and with regularization images with the classic method.

I hope we can find a solution.

Keep up the good work, this technology is evolving very fast!

[TheLastBen]

how many total steps ?

[bach777]

1500, fp16 disabled.

[TheLastBen]

do 1500x4, that's 6000k steps

[TheLastBen]

and keep the fp16 checked

[bach777]

I will try another training and let you know what my results were.

[TheLastBen]

you can retrain (continue training) a previously trained model, use the same images as before

[rodrigoscoelho]

you can retrain (continue training) a previously trained model, use the same images as before

But can you do this after losing the collab files, using only the gdrive ones? I'm lost trying to find out how to resume from the intermediate checkpoints I have saved in the google drive. Is is possible? Or you should have the collab always open? If you close it, you lose the resume possibility?

Thanks and congratulations for the incredible work @TheLastBen

[TheLastBen]

if you have a final model in the sessions folder, you can simply load the session using its name and it will load the previous model without having to run the "model download" cell.

if you want to use an intermediate as the initial model, simply copy its path using the colab explorer and paste it in the "model download" cell "path_to_trained_model", don't forget to check the box "resume training.

[rodrigoscoelho]

thanks. Understood.

Em dom., 30 de out. de 2022 às 10:38, Ben @.***> escreveu:

if you have a final model in the sessions folder, you can simply load the session using its name and it will load the previous model without having to run the "model download" cell.

if you want to use an intermediate as the initial model, simply copy its path using the colab explorer and paste it in the "model download", don't forget to check the box "resume training.

— Reply to this email directly, view it on GitHub https://github.com/TheLastBen/fast-stable-diffusion/issues/251#issuecomment-1296261267, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABUXTGSPYE4X5YPKUHMJ4JLWFZ25BANCNFSM6AAAAAARPBELT4 . You are receiving this because you commented.Message ID: @.***>