On: DeepDreamVideo

[andybak]

I'm really curious to see how much the original training material affects these images.

Is everyone using the same source? There's a lot of doge in there along with that bird that always pops up. Why are faces so prominent? Is that an artefact of the training data or inherant in the algorithm. I would guess the former.

[Houshalter → andybak]

It was trained on Imagenet. 30-40% of Imagenet is just distinguishing different breeds of dogs. And the remainder is mostly other types of animals and random objects.

[sp332 → andybak]

All the versions of this I've seen so far are using the model from Google's github page https://github.com/google/deepdream