Just as Atari and Go are good test environments for reinforcement learning, there are some nice test environments for generative art, the 12 days of Christmas being one of them. The test suite is best appreciated by opening each picture and singing the words! Various algorithms can demonstrate their properties and representational problem solving strategies by attempting to depict the scenes described. A well known story is also useful as the viewer has a stronger prior about what should be seen, and so a more abstract mark can begin to be more easily interpreted, birds being particularly forgiving in this respect it seems.
Here is my attempt to make the 12 Days of Christmas out of only 50 or 80 leaves using the Arnheim variants developed by Piotr Mirowski, Dylan Banarse, myself, et al. Some are more successful than others, but I am blown away by how much veridicality can be achieved (if you squint) with only 50 transparent leaves. A collage is greater the fewer its parts I think, and with some of the more complex scenes I found it necessary to use more parts.
Some particularly ingenuous decisions are made with the swans in the lake, and the beaks of birds, that I think are supra-human in some cases, to the extent I don't think many humans would have thought of making the collage like that, certainly I would not have .I have thrown away several instances for each prompt (up to 20) and in some cases I was not sure which was better. An alternative swan is found at the end, which I think is spectacular in its own way.
[The End, Merry Christmas, wishing you all a lockdown free holiday!]
The 13th Day of Christmas goes to the watery and light swans in this picture, best seen from far away printed very large.
Other test suits of this nature include "The14 Stations of the Cross" a subject much more popular in art, "The Alphabet", and at a stretch the "The Periodic Table", The illustration of children's books is another potential use for these algorithms. The capacity to choose ones primitives has been tremendously freeing compared to other works that use GANs and guided diffusion (at least for me as an artist). Producing these works has really been a very fulfilling collaboration, full of surprises and things learned about initial conditions, hyperparameters, etc.. It is nice to test one variant of the algorithm fully, pushing it to its limit, before moving onto other algorithms as one gets bored.