Evolving AI Art

Emergent Garden
Oct 29, 2022
10 notes
10 Notes in this Video

Pick Breeder: Evolutionary Art Through Selection and Mutation

PickBreeder EvolutionaryArt SelectionMutation InteractiveEvolution
00:28

Pick breeder, dating back to the 1980s with a 2008 experiment inspiring modern versions, allows communities to evolve images through selection and mutation cycles, breeding pictures like organisms.

Stable Diffusion's Open-Source Flexibility: Image Seeds and Prompt Mutations

StableDiffusion OpenSource ImageToImage FlexibleEvolution
01:55

Stable Diffusion is an open-source text-to-image model that allows running on personal GPUs, enabling capabilities unavailable in proprietary services like mid-journey.

Interpolation: Automated Smooth Sampling Through Latent Space

Interpolation LatentSpaceSampling AutomatedEvolution SmoothTransitions
02:19

Interpolation methods allow generative models to run autonomously without human selection, resulting in smooth videos of evolution through latent space.

Original PickBreeder 2008: Untrained Neural Networks for Open-Ended Evolution

PickBreeder UntrainedNetworks CommunityEvolution OpenEndedCreativity
02:29

The experiment that inspired modern evolutionary art was PickBreeder from 2008, a community effort where people selected favorite images from populations that were then mutated and crossbred into the next generation, still playable today.

Image Space: The Incomprehensibly Vast Museum of All Possible Pictures

ImageSpace HighDimensionalSpace CombinatorialExplosion VectorSpace
02:53

Image space is a literal mathematical space containing every conceivable image as a unique point, visited virtually through the Library of Babel’s endless hallways of randomness.

Text-to-Image Models as Generative Search Engines: Creation Meets Discovery

GenerativeModels SearchEngines LanguageNavigation CreativeSearch
03:44

Text-to-image models like mid-journey and stable diffusion grant users an extremely powerful navigational tool—language—to explore the otherwise unsearchable vastness of image space.

Latent Space: Navigation Coordinates Mapping to Image Space

LatentSpace LatentVectors GenerativeModels NavigationalCoordinates
04:06

Text-to-image models use latent vectors—sets of random numbers forming coordinates in latent space—as essential inputs alongside text prompts.

Prompts and Latent Vectors as Genetic Code for Generated Images

GenotypePhentotype EvolutionarySearch ParameterOptimization SelectionMutation
04:20

The input prompt and latent values function as the genes of the output image in text-to-image generative models, enabling evolutionary search strategies.

Latent Explorers: Sailing Uncharted Seas Guided by Interestingness

LatentExploration NavigationMetaphor CompassInterestingness CreativeDiscovery
05:01

Generative models serve as both our map and our ship on which we sail the uncharted latent seas, guided by the compass of interestingness, seeking undiscovered islands of beautiful images.