Really cool, especially the morphing stuff which is amazingly convincing. On the other hand the drawing stuff, I just still can't get over the feeling that it's very similar to looking up best matches in a database, perhaps slightly morphed to better match the input. I guess that's more or less what it is..
Not really - it is sampling from a manifold. The manifold is constructed to fit over the examples given to it - which are the "matches" you are referring to. So while the manifold is constructed from examples, one can sample from anywhere on the manifold. This is very likely what an artist is doing when he draws something from memory.
The morphing is impressive. I wonder if something similar can morph between faces. It may even be possible to morph faces of parents and get something similar to the children of the parents, especially if that's set as the objective. Looking forward to release of the code.
Thank you guys for the interest. The interactive image generation interface as well as trained models are available at https://github.com/junyanz/iGAN. It is still under active development and let me know if you find bugs or would like to suggest new features.
10 comments
[ 6.7 ms ] story [ 33.2 ms ] threadReally cool, especially the morphing stuff which is amazingly convincing. On the other hand the drawing stuff, I just still can't get over the feeling that it's very similar to looking up best matches in a database, perhaps slightly morphed to better match the input. I guess that's more or less what it is..