When will the code be released?

[Devetec]

What is the timetable for the code being released?

[hubert0527]

Thanks for your interest! The codes are all set, they will be released upon the paper is accepted somewhere, so it is highly depending on the reviewing process.

[MoemaMike]

will the code accept with any pre-existing models? I have trained a few stylegan2-ada and stylegan-ada-pytorch models. It would be great if i could upscale them

[hubert0527]

I'm afraid it can't, the architecture is different.

[MoemaMike]

i see. Thanks for the replies. My objective is artistic, therefore high resolution is very important to me. I use my original high res photographs 16mp to 36mp to construct a dataset. With stylegan-2-ada I downsampled these images to create 1024x1024px datasets. Results were good but resolution is less than desirable. I can upscale them using things lin ON1 Resize or the newly released Adobe Super Resolution but I am interested in best results possible.

In an attempt to get better resolution generated images I am in the process of using a dataset of my photographs at 2048x2048 res with stylegan2-ada-pytorch but because of resource constraints on Colab Pro i have to restrict the dataset to 1000-1500 images as the datasets are over 20-30gb in size. And naturally the processing time per tick is slower by half.

My question for you is that with Infinity GAN what would you recommend to someone like myself who has the freedom to create arbitrarily large datasets with arbitrary resolutions? Effectively what is the sweet spot to maximize the quality of the output?

One orthoganal idea i have is the possibility of creating dataset(s) out of the generated images from my stylegan2-ada models and feeding those outputs as input datasets to Infinity GAN. What do you think of that idea?

[hubert0527]

My question for you is that with Infinity GAN what would you recommend to someone like myself who has the freedom to create arbitrarily large datasets with arbitrary resolutions? Effectively what is the sweet spot to maximize the quality of the output?

Our method is still limited to certain datasets that do not have very strong and long-term structural relationships. It works well on landscape or scenery images, but unlikely to work on object-centric datasets such as ImageNet or CelebA. We haven't carefully investigated the relationship between the training patch size and the full image size.

One orthoganal idea i have is the possibility of creating dataset(s) out of the generated images from my stylegan2-ada models and feeding those outputs as input datasets to Infinity GAN. What do you think of that idea?

I'm not very sure about this. Ideally, you can directly train InfinityGAN on the dataset. All GANs are trying to match the distribution of the synthetic dataset and the real dataset, so StyleGAN2-Ada and InfinityGAN are learning to match the exact same distribution (just with different strategies), and StyleGAN2-Ada does not seem to introduce any additional information here.

[harrytrinh2]

Great paper. Hope that it'll be coming soon :)

[mikigom]

still waiting for release 😢

[Harry-Loud]

God, I would love to try this out ASAP! This AI does exactly what I require.Is there any chance I could get access to the code privately?If not, when do you think this will be accepted?

[ucalyptus2]

@FramingApp after iclr 22 decisions

[veryhigh]

What is the timetable for the code being released?