Skip to content

Latest commit

 

History

History
35 lines (24 loc) · 1.17 KB

README.md

File metadata and controls

35 lines (24 loc) · 1.17 KB

tiled-dreamer

Deepdream for any size image.

I got annoyed with artists keeping their any-sized deep dream code closed source, and making a big deal about it.

So made my own in a few hours - peow!

Make sure you have caffe installed and it's python modules in your PYTHONPATH

pip install -r requirements
python tiled_dreamer.py -i my_large_input.jpg -o my_large_output.jpg -t inception_4e/3x3

Also of interest might be the --explore option, which will iterate through the layers and try dreaming on each (while filtering our the dummy "split" layers Caffe creates).

How it works

To deepdream at any size:

  • split into tiles, but overlap the tiles by N pixels(here N is hardcoded to 32), this ensures the edge pixels have the required context
  • jitter each tile in the same direction. This means saving the jitter of the first tiled moving all tiles by the same amount.
  • combine the tiles, by discarding the edges (half of the overlap amount), also deal with boundary conditions
  • when you move up an octave, split into new tile sizes.

License

Fucking Apache because code adapted from the original deep dream ipython notebook