Midjourney expands into AI creative writing with new diversity techniques

Midjourney‘s AI research expands beyond image generation into creative text generation, tackling a fundamental challenge in LLM outputs. The company has partnered with NYU researchers to develop new techniques that enhance diversity in AI-generated text while maintaining quality, marking an important advancement in making AI writing more imaginative and less predictable.

The big picture: Midjourney, with its massive user base of nearly 20 million on Discord alone, is venturing beyond its core image generation business by tackling limitations in AI-generated creative writing.

• The company collaborated with NYU researchers to develop two new techniques called Diversified Direct Preference Optimization (DDPO) and Diversified Odds Ratio Preference Optimization (DORPO).
• These methods specifically address the tendency of large language models to produce homogeneous, predictable text when asked to generate creative content.

The technical innovation: The research introduces “deviation” as a training metric to guide language models toward producing more varied and creative outputs.

• Unlike existing diversity-promoting techniques that operate only at inference time, these new methods integrate diversity considerations directly into the model’s training process.
• The researchers published their findings in a paper on Hugging Face, complete with implementation details available on GitHub.

Key findings: Tests showed that DDPO significantly outperformed standard approaches in balancing quality with diversity.

• Llama-3.1-8B models trained with DDPO achieved the best balance of output quality and creativity.
• Remarkably, even with reduced dataset sizes, DDPO-trained models maintained their ability to generate diverse outputs.

Why this matters: Creative writing fundamentally differs from factual or code generation tasks because it benefits from multiple valid approaches rather than a single correct answer.

• Standard LLM training methods often prioritize user preference over originality, resulting in safe but repetitive content.
• By addressing this limitation, Midjourney’s research could lead to AI systems that better support human creativity rather than merely producing predictable, templated content.

Behind the numbers: Midjourney’s expansion into LLM research comes after the company revealed plans to develop its own computing and AI hardware in late summer 2024.

• With “nearly 20 million users on its Discord channel” according to third-party trackers, Midjourney has the scale and user base to potentially apply these creative writing improvements across multiple generative AI domains.