[Feature Request]: Implementation of NovelAI Aspect Ratio Bucketing

[Lukxie]

trafficstars

Is there an existing issue for this?

• [X] I have searched the existing issues and checked the recent builds/commits

What would your feature do ?

Current training method usually requires a set of images at a 1:1 aspect ratio to achieve somewhat decent results.

This feature should ease the training process and also prevent the loss of quality and info the cropping causes.

Proposed workflow

Option for Aspect Ratio Bucketing in the training/train tab

Additional information

https://github.com/NovelAI/novelai-aspect-ratio-bucketing

[timntorres]

Doesn’t that implementation allocate different buckets to different GPUs? It doesn’t seem too relevant for training on 1 GPU, at least not without some tweaking...

Besides, the type of training this is meant for lies far beyond any consumer’s use case, no?