All images have the same structure as in the paper.

  - row 1: input image
  - row 2: output from filter (x')
  - row 3: output from generator (x'', s'=0)
  - row 4: output from generator (x'', s'=1)

The folders contain:
  
  - 0_smiling_high_res  : smiling, 224 by 224 pixel image results
  - 1_smiling_eps_0.001 : 64 by 64 pixel randomly chosen images
  - 1_smiling_eps_0.005 : 64 by 64 pixel randomly chosen images
  - 1_smiling_eps_0.01  : 64 by 64 pixel randomly chosen images
  - 1_smiling_eps_0.05  : 64 by 64 pixel randomly chosen images
  - 2_gender_high_res   : gender, 224 by 224 pixel image results

This is to show additional qualitative results, and also the effect of the
distortion budgets (epsilon) on the qualitative results. The reason images are
smaller is computational limitations.
