any image from validate set can be fed into the trained IC-GAN model? how?

[fido20160817]

Here is my idea: generated feature for k-means instances in training data; given any image, find the closest center; generate images conditioned on this center feature.

Am I right?

[ArantxaCasanova]

Hello,
the 1000 instance features provided in this repository, chosen with K-means on the training ImageNet dataset, are meant to be used if one wants to sample from the ImageNet distribution, without having access to the full dataset.
In practise, the model can be conditioned on features from any input image, performing zero-shot transfer to potentially other data distributions (note: the closer those images are to the ImageNet distribution, the better the generated images will be). In the Colab notebook we give the option to upload an image to condition the model with.