Thread by @AndresKarjus

New preprint:
"Compression ensembles quantify aesthetic complexity and the evolution of visual art" 🎨🖼️📈
by @AndresKarjus @schichmax @sebastianahnert @mcanet @tillmannohm
🔓arxiv.org/abs/2205.10271
1/🧵

2/7 How does it work? Take an image, run it through a large n of visual filters (blur, gray, edges, etc), compress the variants, record file sizes. Yields a vector where each value expresses some aspect of visual aesthetics, resulting in an explainable vector space. Simple&fast!

3/7 But does it... work? Test against human vis complexity ratings: correlates very well &outperforms plain compression. Also designed downstream tasks such as author identification: also works quite well. Again, no #deeplearning or pretrained models here, just image compression!

4/7 So it's cognitively valid & captures visual family resemblance. Let's do quant #arthistory on 70k artworks from last 500y. The space is high dim, but we can boil it down to a few interesting components using e.g. PCA. Trends correspond to narratives in art history quite well.

5/7 How about something contemporary? We mined the first few months of #hicetnunc for #NFT art and investigate its dynamics over time. We also show these vis #aesthetics estimates describe about 8% of variance in a linear model predicting NFT #sales prices.

6/7 We also look into artistic career trajectories, by quantifying "temporal resemblance", how distant in time a given artist's works are to their nearest artist neighbors in vector space. The model provides interesting insights, disambiguating careers with distinct trajectories.

7/7. Others have used img compression to probe visual complexity before (in veery dif fields!), but our approach goes a step beyond, constructing an efficient model that works better than compression-only & where each dimension remains explainable via the transforms (unlike #ML).