Midjourney releases V4 in alpha

Background

Midjourney's product turns a written prompt into images: the user describes a scene, a style or an object, and the model returns pictures meant to match the description. The company numbers its image models by generation, and a new generation typically starts as an alpha — an early release that users can try while development continues.

The release

On 5 November 2022, Midjourney released V4 in alpha.^[1] It was the fourth numbered generation of the company's image model to reach users, and it arrived as a test version rather than as a finished product.

A change in the compute stack

The most concrete technical shift came in training. Starting with V4, Midjourney's models were trained on Google TPUs.^[1][2] A TPU, or tensor processing unit, is a chip built specifically for the kind of arithmetic that machine-learning models do — large batches of matrix multiplications — rather than for general-purpose computing. Training a text-to-image model means showing it very large numbers of image–text pairs and repeatedly adjusting its parameters so that its outputs move closer to the pairing it was given; that loop is compute-hungry, and the hardware it runs on shapes how quickly and at what scale a model can be trained.

Google Cloud, which supplied that hardware, said that Midjourney rendered images on Google Cloud GPU virtual machines running NVIDIA GPUs, while TPUs were used for training.^[2] The split is a common one in AI systems. Training is a long, throughput-oriented job that runs for as long as it takes; generating an image for a waiting user is a short job where latency matters more. Graphics processors, designed for rendering and later found to be well suited to parallel arithmetic, became the standard hardware for deep learning, while special-purpose accelerators such as TPUs were built to take on the heaviest training workloads.

Reception

The Decoder described the jump from V3 to V4 as huge, with images that were more appealing and detailed and that matched the input prompts better.^[3] Prompt adherence is a recurring weak point for text-to-image models — an image can be attractive and still get the subject, the count or the arrangement wrong — so a reported improvement there is the kind of change users notice immediately, because it shows up in the most ordinary use of the tool: fewer rewrites of the same prompt to get something close to what was intended.

Midjourney also collected community ratings during the V4 alpha and said the process further improved the quality of the final release.^[3] Ratings of this kind usually work by showing users pairs of generated images and asking which one is better. The resulting preference data is a common way to tune a model toward what people actually choose, rather than toward what an automatic metric alone rewards, which puts a large number of users' judgements into the development loop.

What followed

The anime-focused Niji model followed on 20 December 2022.^[1] Rather than covering the broad range of subjects a general model is asked for, it concentrated on a particular visual style, which marked the beginning of the product line splitting by style and intended use.