Better LoRA by organizing datasets by body part?

[PetrHuu]

First, thank you to @bmaltais to keep this project.

Consider using batch size 8 for training.

Dataset:

32 Headshots, aspect ratio 1:1
24 Upper body shots, aspect ratio 4:3
8 Full body shots, aspect ratio 3:2

Idea:

The bucketing should make 3 buckets by the aspect ratios. Because of batch size 8, similar body parts should be trained in the same batch.

Will it help the likeness of the person in the LoRa model?

Thank you.