Given a fixed context, we sample two latent codes from the prior distribution and linearly interpolate between them. The strokes smoothly transition from the start to the end point, focusing on the same object.
In this section, we compare stroke sequences produced by different methods. We set the sequence length K=24. We can notice that I-Paint tends to predict sequences of strokes that are distributed over the subject and color consistent, which resembles the painting style of the datasets.
In this section, we show sampled video sequences from the Oxford-IIIT Pet INP and ADE 20K Outdoors INP datasets. We show continuations of 24 strokes, highlighted in red.
Oxford-IIIT Pet INP
ADE 20K Outdoors INP