[FEATURE REQUEST] text-to-image finetuning with aspect bucketed data

[bghira]

Feature Request

I would like for autotrain to be able to train on aspect-bucketed datasets.

Motivation

When training text-to-image models, it is beneficial to have data bucketed into a few aspect ratio sizes.

The images can be random-cropped into these buckets, preferably with an optional downsample size before cropping so that we don't lose much scene integrity (eg. you wouldn't want to go from 4000x4000 to 1024x1024, maybe you'll just get an image of a wall)

Additional Context

Currently, only "dreambooth" finetuning is supported, and it seems like maybe the model config comes from the hub repo directories.