My immediate thought is that if automation is really needed for the preservation of traditional patterns, then the right tool for the job is not unreliable, only probably approximately correct, generative AI approaches like GANs and Diffusion models, but Good, Old-Fashioned, symbolic AI approaches widely used today for procedural asset generation in the games industry like L-Systems or texture synthesis algorithms and Wave Function Collapse.
Wave Function Collapse in particular is itself based on convolutions, like CNNs, to preserve translational invariance, but it also preserves equivariance by means of a constraint satisfaction procedure. It's cool stuff and dirt cheap compared to training a CNN from scratch*.
Wave Function Collapse in particular is itself based on convolutions, like CNNs, to preserve translational invariance, but it also preserves equivariance by means of a constraint satisfaction procedure. It's cool stuff and dirt cheap compared to training a CNN from scratch*.
_______________
* On anything but ImageNet/CIFAR 10.