1. Conditional generation

• Conditional Generation

  • Ouputs from a asked class
  • Labeled datasets
  • Inputs Noise vector + One-hot vector
    • Concatenated

• Unconditional Generation

  • Ouputs from a random class
  • Unlabeled datasets
  • Inputs Noise vector

1-2. Inputs

• Generator input
  • Noise concatenated with one-hot vector

• Discriminator input
  • Image labeled with one-hot vector
  • Many ways of assigning a label to an image e.g one hot matrix

2. Controllable Generation

Tweaking the input noise vector to get different features on the output

• Controllable Generation

  • Examples with the features that you want
  • Training dataset doesn’t need to be labeled
  • Manipulating z vector input

• Conditional Generation

  • Examples from the classes you want
  • Training dataset needs to be labeled
  • Appending class vector to the input

2-1. Vector Algebra in the Z-space

Interpolation Using the Z-space

Interpolating between two input vectors in z-space allows us to get the ‘gradient’ between to images

Z-space and Controllable Generation

The goal is to find the directions for different features we care about

2-2. Challenges with Controllable Generation

Feature Correlation

• Difficult to control a specific feature without modifying others
• When different features have high correlation in the data set

Z-space Entanglement