Optimizing Training Settings for Fast DreamBooth

[sahilsingh2402]

Hello, I'm seeking advice on optimizing the training settings for Fast DreamBooth, specifically for subject/face recognition. Several months ago, I achieved satisfactory results with the default settings, but lately, when using the same settings and model for training, I've been encountering overfitting issues.

I reviewed the discussions, but I couldn't locate any relevant information pertaining to the most recent notebook. All the suggestions and recommendations seem to be tailored to earlier notebook versions.

I currently have a dataset comprising 15 images of the subject. Could you kindly recommend the ideal number of training steps and learning rates for the UNet and TextEncoder in this context? Your insights and guidance would be greatly appreciated. Thank you.