Trained resolution and memory requirements

[morgankohler]

What resolution were these models trained at? I tried 1920x1080 images and this results in OOM even on 80gb A100. Is this far higher than expected resolution?

[kai422]

Hi morgankohler,

I think for training, 1920x1080 images will likely go OOM. We use 64x64x16 (WxHxT) as input for training, where the gt is of size 256x256x16. 1920x1080 will results in 500 times more memory.

You can crop the image into smaller patches and use them for training.

Regards,
Kai

[Platorius]

Hmm, but doesn't it destroy the whole point of VSR, if you tile the images?

[Manoa1911]

don't expect a answer, he thinks he is smart: if he gives you a truth response to this question he will have to admit that his soft is shit, if he lies it will be easily cut through by facts that prove he lies, so he thinks that if he says nothing it makes him smart

[Ko-Lani]

I have to say, that last comment is unnecessarily mean and unproductive. The OOM issue isn't unique to this model; it's a fundamental challenge for most complex vision models due to hardware limitations. It's unfair to single out one project as if it's uniquely flawed.

Instead of one-sided attacks, we should be discussing constructive solutions. For example, exploring VAE-based approaches to handle high-dimensional inputs would be a far more valuable conversation for everyone. This kind of slander is not appropriate here.