Midjourney V3

Midjourney V3 is the third model version of Midjourney, a system that generates images from text prompts. It was released on 25 July 2022, shortly after Midjourney's open beta on 12 July 2022 ^[1]. In the project's line of versions it marks a sharp jump in image quality, and it is the version behind one of the most widely discussed AI-made pictures of that year.

How people reached it

Users worked with it through a Discord bot: a prompt was sent in a channel, and the generated images surfaced for all users in that channel in about a minute ^[3]. That detail matters for understanding how the model developed, because the tool was used in public, in shared spaces, with many people watching results appear at the same moment.

What changed

According to Midjourney's founder David Holz, V3 added a feedback loop built on user activity, and he credited that loop with a large part of the quality improvement ^[2]. He described the gain as far bigger than one would expect, noting that the team did not put any more art into the training data; the improvement instead came from data about which images users liked and how they used the tool ^[3]. In other words, the raw material of the training — pictures — did not change much. What changed was the information about what people actually preferred.

A feedback loop, in this setting, means the system keeps track of how people respond to what it produces and uses those responses to adjust the model. Instead of only showing a model more examples of images, you show it which of its own outputs people chose, kept, or worked with, and let that steer it. For image models in general, output quality depends not only on how many images the model has seen but on how well it is tuned toward what people want; gathering signals from real use is one way to do that tuning.

What it produced

In August 2022, an image made in this period, titled Théâtre D'opéra Spatial, won the digital-art prize at the Colorado State Fair ^[1]. The award drew attention well beyond AI research circles, because it placed a generated image in a competition meant for human artists.

Where it sits in the line

V3 was followed quickly by test models: --beta on 22 August 2022 and test/testp on 28 August 2022 ^[1]. Its successor, V4, arrived in November 2022 ^[1]. The short gap between the two shows how fast the versions were moving at that stage: V3 was not a settled end point but a step that was being tested and revised within weeks.

Why it matters

V3 is the version where Midjourney's output became good enough to take an art prize ^[1], and where the team's method shifted toward learning from users rather than only from more images ^[2][3]. Both of those points carried forward into later versions: the public, shared way of using the tool, and the habit of tuning the model on how people actually behaved with it.