Question: Does using a higher resolution training set, such as 2048 or above, lead to better performance?

[houmo03]

This is for bugs only

Does using a higher resolution training set, such as 2048 or above, lead to better performance?

The hardware has 48G of video memory, which can support the training of higher resolution material sets, such as 1536, 2048, 2560 and so on, and the original material details of the training set are also sufficient to support it.

This change, compared to the training results of 1024 resolution, will it bring corresponding benefits, or will it have a counterproductive effect?

Thank you

[NBSTpeterhill]

that depends on many things, but FLUX itself is 12B*[10241024] , so I don't think 20[2048*2048] datasets will improve details. The quality of datasets don't only means [resolution], but also others things, like tags, lr, optimazer and so on and the contents inside images.

[houmo03]

that depends on many things, but FLUX itself is 12B*[1024_1024] , so I don't think 20_[2048*2048] datasets will improve details. The quality of datasets don't only means [resolution], but also others things, like tags, lr, optimazer and so on and the contents inside images.

Thank you for your response. If it is difficult to obtain better benefits from flux, I will switch back to 1024 for training, and cut the high-resolution images into multiple images to use. Looking forward to a better upscaling process to assemble the details in the future.