Data
Z
W
W+
Overall I tuned the lr=0.5 to speed things up without noticable dip in quality.
In the projection, I actually found W to be slightly more appealing, because it seemed to be slightly more lenient and thus have a little less artifacts, particularly in the last row:
Data
Z
W
W+
Interpolation was fun, the non-linearities in the transition despite the linear interpolation of the latent vector is pretty cool.
Data A
Projection A
Data B
Projection B
Interpolation
I tried many different ratios of perceptual loss and l1 loss. I enjoyed the more wacky results, so there's more of those.
The denser drawings put more constraints on the image, which makes the cats look more like a solid block of color. I prefer the sparser drawings (see the next section for my own drawings), which give more texture as opposed to blocky regions.
Overweighting the l1 loss makes the result especially close to the input, which can result in bright colors, but sacrifices catness.
I aimed for bright colors with furry textures by using about equal l1 and perceptual loss, while drawing with thin lines.
Drawing
Result
The background of this website is a mix of drawing and interpolating, here are the drawings:
Drawing
