Encoding interpolation varies latent representations between examples and decodes the intermediate representations into a smooth sequence of outputs. It turns a trained autoencoder or variational autoencoder into a tool for exploring gradual changes in generated images. section 7.0
A typical workflow is:
- Encode example inputs into the model’s lower-dimensional latent space. section 7.0
- Vary or interpolate between the encoded representations. section 7.0
- Decode the intermediate representations to generate new outputs. section 7.0
Glasses transition
A trained variational autoencoder can generate a series of face images transitioning from glasses with dark frames to glasses with light frames. section 7.0
Figure 7.2 Generating a series of images that transition from glasses with dark frames to those with light frames
Browse the visual sequence showing a gradual change from dark-framed to light-framed glasses. section 7.0
