Credit: VentureBeat made with Midjourney

Midjourney is best known as one of the leading AI image generators — with nearly 20 million users on its Discord channel, according to third-party trackers, and presumably more atop that on its website — but its ambitions are beginning to expand.

Following the news in late summer 2024 that it was building its own computing and AI hardware, the company this week released a new research paper alongside machine learning experts at New York University (NYU) on training text-based large language models (LLMs) such as Meta’s open source Llama and Mistral’s eponymous source models to write more creatively.

The collaboration, documented in a new research paper published on AI code community Hugging Face, introduces two new technieques — Diversified Direct Preference Optimization (DDPO) and Diversified Odds Ratio Preference Optimization (DORPO)— designed to expand the range of possible outputs while maintaining coherence and readability.

For a company that is best known for its diffusion AI image generating models, Midjourney’s new approach to rethinking creativity in text-based LLMs shows that it is not limiting its ambitions to visuals, and that, a picture may not actually be worth a thousand words.

Could a...