Can the algorithm for automatic scaling of ARB buckets be specified?

Currently, there is already an option to enable or disable the automatic upscaling of images that are smaller than the training resolution to match the training resolution. Could we add another option that allows us to choose the scaling algorithm for either upscaling or downscaling? For example, when training pixel art, I would like to specify that the script uses the nearest neighbor method, while for training regular images, it could use Lanczos or a similar algorithm. Alternatively, the script could automatically detect whether the corresponding txt annotation file contains any pixel-related tags—if it does, it would use the nearest neighbor method, and if not, it would default to Lanczos. This would not only save time by eliminating manual processing of the training set but could also improve the training results, since the scaling algorithm for pixel images shouldn’t be the same as that for regular images, right?

Start by locating the existing automatic upscaling option and the image-resizing path used when building ARB buckets. Clarify whether completion means a user-selectable algorithm, pixel-tag detection, or both, then define how the chosen behavior should apply to upscaling and downscaling.

Written by the indexing model from the issue text.