[johndough]

You'll probably hit memory limits if you go much beyond 512x512-sized holes.

Additionally, computation times grow quickly with higher resolution and you already need a high end GPU for this resolution to get a reasonably interactive response time.

You'll also need a favorably licensed pretrained model or a few 10000 training images and masks.

So all in all, I can't see any deal breakers, but I'd probably still use PatchMatch instead.

[rhcom2]

For reference the GPU they're using for this paper is the NVIDIA V100 GPU, a datacenter GPU costing $8,000.

[penagwin]

To be fair, while V100 perform very very well for machine learning, you can buy almost a dozen 1080ti's or a few titans (whatever the current one is), which would certainly be much faster.

They say they used V100 but not how many, if they needed a large number then nevermind.

[rhcom2]

The paper says they only ran it on a single V100, I was expecting multiple GPUs as well.

[hanrelan]

Since they programmatically generate the masks you wouldn't need those, just the set of training images. So it wouldn't be too hard to find since you're not looking for paired images, just a bunch of images of faces/landscapes/whatever you're trying to inpaint.