You can either generate an artwork from a single prompt (classic method) or create a diverse population of artworks. In the population method, initial individuals are generated from descriptions or existing artworks. The algorithm selects the most novel artworks based on embedding distances, crosses their descriptions using an LLM, and applies random mutations (style, technique, color, content).

Daily, Genmuse identifies popular niches from user interactions, evolving these niches into new artwork generations.

Essentially, it’s novelty search guided by human feedback.

Main model: Flux 1.1 Pro Ultra Embeddings: Voyager Agent includes a small library for post-processing.

Feedback are very welcome!