Midjourney News 11/05/2022 AI Rating: Medium

Midjourney V4 Released: A New Architecture Brings a Major Leap in Image Quality

#Midjourney#Image Generation AI#AI History

Midjourney released “V4” as an alpha version — a new model trained from scratch on the company’s newly built AI supercluster. Developed on a different codebase and architecture from V3 and earlier, it marked Midjourney’s first major generational leap.

Details

  • Development time: A completely new architecture, trained for more than nine months on Midjourney’s new AI supercluster
  • Quality gains: Expanded knowledge of people, creatures, objects, and places, along with improved handling of complex prompts involving multiple subjects
  • Fine detail: Reduced the detail artifacts that had plagued earlier models, enabling more realistic texture rendering
  • What followed: Became the default model on December 20, 2022, and remained the standard until being replaced by V5 on March 30, 2023

What happened next

V4 was the first result of Midjourney establishing its own infrastructure for training large models from scratch, and it became the starting point for the rapid evolution that followed through V5, V6, and V7. Around the same time, development of the anime-focused “niji” model was also underway, and Midjourney would go on to see multiple generational shifts in just over a year.