For text2image LORA-training, it is beneficial to have images with equal dimensions, because multiple images can be combined in one batch. Scaling down images is no problem, but changing aspect ratio is! I would therefore want to restrict aspect ratios of RectangleLabels right from the beginning. The rationale is, that the human is in a better position to decide on relevant detail than any automatic cropping function down the pipeline.
I did not find anything on aspect ratios in the docs, so I assume it would be a feature request?
Allowing multiple aspect ratios per Rectangle and “snapping” to the closest one would be perfect. Only allowing one fixed aspect ratio would also be helpful, there’s always the workaround to have mutliple Labels, one for each A/R.
Would anyone else find that useful? Comments welcome.
Greets, Andreas