This is an odd one. It's an embedding trained on 20 images of a day's Page 1 of The New York Times.

Honestly, I expected the outputs here to be more consistent, given that many elements of the layout are identical from day to day (most obviously the top nameplate and ears). But the variations are actually rather striking, bringing in a bolder use of imagery that's more associated with certain European or Latin American newspapers (or, more often, magazines). It seems to really like abstract oceanscapes and glaciers.

It's not especially responsive to prompting, but I can imagine it being useful for layout inspiration; certainly it produces more visual interest than most newspaper front pages.

Because of the Times' broadsheet layout, this TI is trained on 512x930 images — 0.004 learning rate, batch size 4, gradient steps 5, 200 total steps.