Midjourney releases V6

Background

Midjourney is an image generation service: a user describes a picture in words and the service returns images. Each numbered version of the underlying model sets what the service can render and how quickly it responds, so a new major version is the main way the product itself changes.

What happened

On 21 December 2023 Midjourney released V6 in alpha ^[1]. The label alpha signals that a version is put in users' hands before it is treated as final, so work on it continues after release. This was not a small adjustment to the previous model: V6 was trained from scratch over nine months ^[1].

The improvement Midjourney emphasised was text rendering — images made with V6 handled written words better than earlier output ^[1]. Letters are a hard case for image models, because each character has to take a precise shape and stay consistent with the rest of the picture rather than dissolving into texture, so a model that draws text more reliably is doing something noticeably harder than drawing a landscape.

V6.1

V6.1 followed on 30 July 2024. It improved coherence — the internal consistency of an image, so that parts of a scene hold together instead of drifting — and made standard jobs about 25% faster ^[2].

The release also added new 2x upscalers, tools that enlarge an image to double its size, and a `--q 2` mode ^[3]. That mode took 25% longer to run and sometimes returned more texture ^[3]. The two settings pull in opposite directions: one is a quality dial that costs time, and the upscalers are a separate step applied to an image after it exists. When V6.1 shipped it was made the default model for all users on Midjourney, so anyone who did not deliberately pick an older version was generating with it ^[3].

The website

In August 2024, alongside V6.1, Midjourney's website interface launched ^[1]. Until then the service had been reachable through other means; the website gave the same model a browser-based front end, so a version release and a change in how people reach the model arrived together.

Why it matters

Two things stand out in this sequence. The first is the decision to train from scratch rather than keep tuning the previous model. Retraining means the capabilities on offer are not bounded by earlier design choices, but it also means a long build — nine months — before anything ships, which is why the result was pushed out as an alpha rather than held back for a finished release ^[1].

The second is how the versions were delivered. V6 arrived in December 2023; the refined V6.1 arrived in July 2024 with measurable speed and coherence gains, extra tools, and a place as the default ^[2][3]. Making a version the default is a quiet but consequential act: it decides what most people actually use, independent of what the release notes say. The `--q 2` mode shows the same logic at a smaller scale — quality is exposed as an explicit trade against time, so users choose how much waiting a picture is worth ^[3].

For the product's history, the pair mark the point where Midjourney's model moved to a rebuilt generation with better in-image text, then to a faster and more consistent iteration of it, delivered to users through a website as well as through the model itself ^[1][2].