Question about the “resolution” parameter.

[GobbleCrow]

I’ve been messing with OneTrainer for a bit and wanted a 2:1 ratio fine tune of sdxl. The image size I came out with was 1408x704 because they’re multiples of 64 and are as close to the image pixel count as 1024x1024 as it gets. The issue is I don’t know the correct settings for this training. Sampling is pretty self explanatory but the “resolution” parameter on the training tab is a bit arbitrary. Does anyone know what settings I should do here?

[O-J1]

1024 * 1024 = 1,048,576 pixels
1408 * 704 = 991,232 pixels

Disclaimer: I may be reading the code wrong.

The resolution you have mentioned seems to be within the pixel budget as far as I know, so there’s no need to modify it. However, I’d suggest being cautious about using a single aspect ratio, as it could lead to potential issues down the line. Of course, it’s up to you, and you may have your own reasons for that choice. 1472*704 also still fits within the budget.

As an aside make sure you read the wiki: https://github.com/Nerogar/OneTrainer/wiki/Training#main-settings If you ever have a need to set a specific resolution you can only provide a single exact resolution in the train tab. Multiple nonspecific resolutions can also be provided.

specific (fake number): 1920x1080
multiple: 1024,1152,1280

Relevant https://github.com/Nerogar/mgds/blob/master/src/mgds/pipelineModules/AspectBucketing.py