Generating more complex scenarios

[superxueyizou]

Dear author,

You have shown good results in generating data distribution like ImageNet, CIFAR-10 and iNaturalist, which are category-based datasets. I wonder whether you have experimented on generating more complex scenarios, such as DIV2K, LSDIR ?

[LTH14]

Thanks for your interest! We choose ImageNet and CIFAR-10 because they are the most common benchmarks for (unconditional) image generation. We also evaluate RCG on iNaturalist, which contains 10000 categories and 2.6M images, to demonstrate its ability in modeling very complex image distributions.

Generating high-resolution images could be an interesting future direction of RCG. However, DIV2K (1000 images) and LSDIR (84,991 images) might be too small to train a good high-resolution generative model. Our future plan is to scale the current model to large-scale unlabeled datasets, such as OpenImages or DFN, which contain images with high resolution and multiple objects.