What's the difference between UNet_Learning_Rate?

[meunumerotim333]

Hello guys,

the earlier version of Dreambooth the main UNet_Learning_Rate was 2e-6. The recommended for that was train 100-200 steps for each picture.

Now the main UNet_Learning_Rate is 1e-5 that now reccomends 65 steps for each picture.

What's the difference between each one? (1e-5 ... 2e-6 ... etc) I haven't find any article about that.

If someone also could explain the importance about Text_Encoder_Training_Steps would be pretty cool to understand how DreamBooth really works.

Thanks :) Hope this discussion helps the community

[cdefghijkl]

2e-6 = 0.000002
1e-5 = 0.00001
1e-5 is higher/bigger means it can be faster than 2e-6 to reach the optimal but have to be more careful with steps amount.

TE Training is to teach/introduce your training object token/filename to the model.

That's my knowledge. Cmiiw

[cutecoelho]

Ohhh now I undersntand :)

What about 1.5? What's the best configs for face training in your experience? 1.5 still have the best models

[TheLastBen]

reduce it at 5e-6 and increase the unet steps to 2000

[cutecoelho]

thank you buddy :) I'll try

[shyt47]

reduce it at 5e-6 and increase the unet steps to 2000

Keeping the text encoder to 350 with 2000 steps?

[TheLastBen]

yes

[meunumerotim333]

Well.. I've tested all the tips, but in my case 2e-6 still the best for train face on model.. because it perfectly fits on the style of the external model.

Does anyone one know what's the unet_learning_rate equivalent of 2e-6 1000 steps?
Because I'm totally dependent of this value.. if sometime the 2e-6 gone I'll be totally lost.. does anyone could give me a tip of an alternative value on a different unet_learning_rate?