[fock]

some questions:

- on which dataset was it trained?

- how big is the model?

and finally: can one compare the result to just picking the closest faces from the training set?

[oacstevens]

It's trained with a mixture of publicly available datasets of faces. The final model is several gigabyes in size, so it's fairly large. Actually that's one of the reasons we've made this tool - to test our infrastructure with larger generative models.

The model is learning the features that make a convincing face, and generating a synthetic face from those (controlled by the segment map), similar to Nvidia's GauGAN.

[fock]

Haven't looked into these, but what is the proposed advantage over just picking an image (with a much smaller model maybe?) from the database?

[oacstevens]

The power in generative models lies in being able to flexibly generate images that belong convincingly to a set (e.g. faces, landscapes), but that are not actually in the input dataset. E.g. you can make images that look like faces, but that don't belong to any real individual.

[fock]

And where is the use of that? (I guess somewhere, we're at deepfakes et al, but this site still seems to advertise the concept)