hyper bottom heavy

I cleaned up the dataset a little, but its pretty much the same as v1 as far as tags go. Just wanted to revive this old lora on a modern model to see how it performs.

Training Details:

• I used Diffusion Pipe for this training, anything not listed below, was default value.

• resolutions = [[768, 512]]

• enable_ar_bucket = true

• micro_batch_size_per_gpu = 2

• gradient_accumulation_steps = 8

• type = 'lora'

• rank = 32

• dropout = 0.1

• type = 'AdamW8bitKahan'

• lr = 1e-4 (probably could have been higher)