Every frame of this video is imaginary: The places don't exist in the real world. And for every frame, the same (random) inputs are given to four different models -- one making bedrooms, one making dining rooms, one making kitchens, and one making living rooms.

Feb 24, 2019 · 6:53 AM UTC

43
1,466
222
4,176
0
This tweet is unavailable
I love everything about this video so much. Thanks to @gwern for the suggestion of using transfer learning -- this is why the models produce things that have many of the same features, down to the pattern of the bed turning into a couch into a counter.
4
1
71
The results of this training are now integrated into thisairbnbdoesnotexist.com; you can now get a collection of photos that include rooms other than bedrooms.
2
35
6
115
If you're interested in seeing how the model changed over time, you can check out my new "Transfer Demo" page, which shows how the computer learned to "see" different room styles over time. thisairbnbdoesnotexist.com/t…
7
43
Replying to @crschmidt
Is there a way to run a photo of an empty room through this?
1
There's not really a way to influence what the model generates: The inputs are just random numbers, and those numbers don't have any particular meaning. It might be possible to search the space and find a room that looks like a photo you have. (That's my project for tonight.)
1
Replying to @crschmidt
Out of curiosity, is there something to prevent the model from outputting images that are similar (enough) to the input data? I.e. is it possible that some of these rooms did actually exist aside from a detail or two?
1
1
It is completely possible that that happens, but there are 4294967296 ^512th power possibilities, so the chances it showed one are pretty low :) But if you *search* the space, you can find one, e.g. this is a picture of a friend, and his representation in the faces model.
2
1
more replies
Replying to @crschmidt
Fascinating that the in-between states look believable also (art least in the video form). Perception is so easily fooled.
1
Realistically, there are no "in-between" states -- every state is just a different set of random numbers! But the fact that smearing between the values produces a compelling blending effect is definitely cool, and wouldn't happen with every model.
5
Load more