How do you train for styles now? #1503
-
Hello, I've been using this colab (which works great by the way!) to train for styles but now I'm stuck. For awhile now, I been looking through previous discussions and the step by step tutorial on how to properly train a model on a style, but with the changes to learning rate, I feel like I need some clarification or correction on a few things to make things better. Also I'm training on a custom model if that makes any differences. Instance: I'm using 40. should I lower it to get the best results? training: I did 4,000 steps with a rate of 1e-5, I think that the rate was the previous default. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 2 replies
-
1e-5 is a really fast learning rate so less training steps required to learn the concept you're training on so 4k steps might be way too high for 40 images, the text encoder is a bit hard to have a set number to train on again 650 is a lot for 40 images but this really depends on if your style is coming through with your token if its not you probably need more training to the text encoder, concept training steps seems like 1000 is a good start for your 1300 images, and I havnt used style training more than once but I'm pretty sure its right to do it after training the unet for a bit as it doesn't pass the token to the textual embedding helping to not overfit the model. Hope this helps and I am by all means no expert so don't expect it to be a perfect rundown haha |
Beta Was this translation helpful? Give feedback.
-
Can you share the link for that step by step tutorial you mentioned on first post? I'm interested to learn how to train styles too. Thanks you in advice |
Beta Was this translation helpful? Give feedback.
1e-5 is a really fast learning rate so less training steps required to learn the concept you're training on so 4k steps might be way too high for 40 images, the text encoder is a bit hard to have a set number to train on again 650 is a lot for 40 images but this really depends on if your style is coming through with your token if its not you probably need more training to the text encoder, concept training steps seems like 1000 is a good start for your 1300 images, and I havnt used style training more than once but I'm pretty sure its right to do it after training the unet for a bit as it doesn't pass the token to the textual embedding helping to not overfit the model. Hope this helps and…