Help! I need your dreambooth tips/lessons learned in training custom models of a person? #1009

juan9999 · 2022-12-13T19:35:20Z

juan9999
Dec 13, 2022

Does anyone have a summary of best practices/lessons learned training with dreambooth on people?

I've been having issues producing a good custom model using dreambooth.

Here is what I've tried

using SD 2.1 768
using ~40 pictures (renamed to a non-existing work) resized to 756X756 (no other people, different expressions, backgrounds, angles, 20X close 6X side, 10 chest up, 6 full body)
setting UNet training to 1000
400 text encoder training steps (I'm not sure what this does)
I skip Concept Images
I don't "tag" photos which I've heard some people do

I then tried:

26 pictures
setting UNet training 101*[number of pictures] (2626)
But first model seems to produce better results

I haven't used tags,

Most training was done with Tesla T4

TheLastBen · 2022-12-13T20:07:23Z

TheLastBen
Dec 13, 2022
Maintainer

The V2 is harder to train than the V1.5, right now I'm working on finding the optimal settings to train on faces, hopefully soon I will find the right combination of steps and leaning rate.
tip: use only 15-20 instance images for better training.

3 replies

gilroff Dec 16, 2022

The V2 is harder to train than the V1.5, right now I'm working on finding the optimal settings to train on faces, hopefully soon I will find the right combination of steps and leaning rate. tip: use only 15-20 instance images for better training.

Won't more instance images in different settings but still keeping only the same face, improve the training? I am still new to this. I dont think i really understand what concept images exactly do. can you explain how instance image training, text encoder and concept image training actually work in a simplified way? Or is there a documentation like that I can check out? Please give me link if there is. Btw thanks for your awesome work <3

TheLastBen Dec 16, 2022
Maintainer

to many instance images will split the number of steps per image and will result in poor quality, 10-20 instance images is enough to capture the face or the style, then the concept images will help the text_encoder widen it's range of output, concept images can be as many as 200-300 to help with diverse concepts.