Midjourney V4

What V4 is

Midjourney V4 is the fourth-generation image generation model from Midjourney, a service that turns short text prompts into pictures. It was released in alpha on 5 November 2022, which means users could try it while the version was still in an early, testing stage^[1]. Text-to-image models like this learn statistical patterns from large collections of images and then produce a new picture guided by the words a user types; the version number marks a new round of that training rather than a change to the interface idea itself.

How it was trained and run

Starting with V4, Midjourney's models were trained on Google TPUs^[1][2]. Training is the stage in which a model's parameters are adjusted over very large image collections; inference is the stage in which the trained model actually produces an image for a user's prompt. According to Google Cloud, V4 was trained on Google Cloud TPUs and inference ran on Google Cloud GPU virtual machines with NVIDIA GPUs^[2].

On 14 March 2023, Google Cloud said Midjourney trains its latest algorithms on Google v4 TPUs with JAX, and the announcement quoted founder David Holz^[2]. Midjourney also cited sustainability as a factor in choosing Google Cloud: it needed a partner that could help curb its energy consumption and carbon emissions^[2].

What changed compared with V3

The Decoder described the jump from V3 to V4 as huge, with more appealing and detailed images that better matched the input prompts^[3]. Midjourney also ran a community rating process for V4, in which users compared and ranked outputs, and the company said this further improved the quality of the final release^[3].

Related releases and successors

On 20 December 2022, Midjourney released Niji, a model focused on anime-style images^[1]. V4's successor, V5, followed in March 2023^[1]. At the time of the March 2023 Google Cloud announcement, Google Cloud said Midjourney's artistic community had more than 11 million members^[2].

Where it sits in Midjourney's development

V4 is the point at which Midjourney's training moved onto Google TPUs, and the version that outside observers treated as a clear break from its predecessor^[1][3]. Three things are documented about it: a large quality gain over V3 in how closely images followed prompts, a training and serving setup split between Google Cloud TPUs and GPU virtual machines, and a community of more than 11 million members around the service by the following spring^[2][3]. The company also tied the choice of infrastructure to energy use, naming sustainability among its reasons^[2]. V5 arrived four months after V4's alpha release, so V4 was a step in a fast sequence of versions rather than a lasting endpoint^[1].