Skip to content

Conversation

@enlyth
Copy link

@enlyth enlyth commented Aug 22, 2025

Sometimes you might be in a situation where you want to train on your original dataset resolutions without resizing, because you have either pre-resized your dataset, or it has specific requirements that resizing would be detrimental to (such as training pixel-art datasets).

This feature adds a preserve_resolutions boolean option to datasets, defaulting to False, which if True, completely disables resizing.

It basically results in one bucket per each unique image resolution in your dataset, regardless of what it is, without resizing.

image image image

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant