Kasucast #11 - Stable Diffusion: How to train DreamBooth for Style

#stablediffusion #characterdesign #conceptart #digitalart #machinelearning #dreambooth #style #hypernetwork #textualinversion #digitalillustration #aiart #style

🎉 Social Media:

Timestamps:
00:00 Introduction
00:44 Examples of DreamBooth Style txt2img
01:26 Start of ShivamShrirao's local install overview
02:10 Small error regarding bf16 and how to resolve it
02:39 Explanation of modified shell script file
04:26 All you need to change is the amount of dataset images you have
05:13 How to change checkpoint weight save intervals
06:16 Example of hypothetical training dataset
07:23 How to convert diffuser to checkpoint
08:07 Trying out the trained and converted model for txt2img
09:02 Txt2img batch generation

[danielkim]

Nice work! I have a question about collecting images. How do you usually collect images for training?

[kotox]

before even starting, thank you <3 :)

[wolfai_]

Nice result. But looks like you did all of it manually? Why not use the dreambooth extension on automatic1111?

[kasukanra]

I just want to mention that I've shared an example of the training script I used in this pastebin:


Even if you don't use the local install for DreamBooth training (ShivamShrirao in WSL2), the variable initialization (training steps, number of class images, learning rate, warm up learning rate, etc.) at 03:43 is still probably helpful as a reference for other methods (DreamBooth AUTOMATIC1111 extension, etc.).

Also, if you're running into errors in WSL2 after install and it keeps giving you errors, try running this script on the command line:

export


Please read the description for more details as well!

[devnull_]

Stupidly long negative prompts seem to make images usually worse or at least less varied IMO.