santa hat
deerdeer nosedeer glow
Sign In

SDXL LoRA vs SDXL DreamBooth Training Results Comparison


The SDXL DreamBooth config and workflows are shared here :

4k+ Preprocessed Ground Truth regularization images are shared here :

Hopefully a tutorial coming to where I will show settings, stay subscribed

Both trained on RTX 3090 TI - 24 GB

Same epoch, same dataset, same repeating, same training settings (except different LR for each one), same prompt and seed

LoRA has xFormers enabled & Rank 32

I have checked LoRA settings multiple times and they are correct. They from my this video :

These below results are from checkpoint 4. Since repeating was 40 that means 160 epochs

Number of used training images are 13, used my ground truth best regularization images as well

Compared checkpoint 3 and 2 as well and results are same. DreamBooth superiority is something else

Hopefully a tutorial for DreamBooth training is coming on SECourses