Hey everyone,
I’ve been wondering why Topaz doesn’t export its PyTorch models to ONNX and TensorRT with larger tile sizes.
It just seems logical that with bigger tiles, we’d get higher quality results and less need for all that stitching and overlap. Plus, for those of us with beefier GPUs, it would be a huge speed boost!
I get keeping the 128 tile size for compatibility, but why not also give us the option to choose models with tile sizes like 256, 512, 1024, or even 2048? If you look at some of the beta models out there, it’s clear this is totally possible and feasible. ( http://models.topazlabs.com/v1/ggnv2-v1-fp32-2048x2048-2x-ox.tz
http://models.topazlabs.com/v1/ggnv2-v1-fp32-2048x2048-2x-ox.tz2
http://models.topazlabs.com/v1/ggnv2-v1-fp32-2048x2048-4x-ox.tz
http://models.topazlabs.com/v1/ggnv2-v1-fp32-2048x2048-4x-ox.tz2
http://models.topazlabs.com/v1/ghqbeta-v1-fp32-2048x2048-2x-ox.tz
http://models.topazlabs.com/v1/ghqbeta-v1-fp32-2048x2048-4x-ox.tz)
So, it really begs the question: Why the focus on such small tile sizes for the official models? Even competitors like Aiarty Image Enhancer support tile sizes over 1000. It would be awesome if you guys could look into this.
Thanks for listening!
Cheers.
