Midjourney releases V5

Midjourney released an alpha version of V5, the latest iteration of its text-to-image generator, followed within a day by a public beta. The update was pitched as a step up in photorealism, coherence and prompt responsiveness over the previous version, with double the effective resolution and a wider range of supported aesthetic styles.

The detail that drew the most attention was hands. Earlier generations of image models, Midjourney included, reliably produced human hands with the wrong number of fingers or joints bent in impossible directions — a flaw well known enough to be a running joke about the technology’s limits. Reviewers testing V5 reported that it rendered hands correctly far more often than not, and several described images that could pass, at a glance, for photographs rather than synthetic output. The improvement was anecdotal rather than benchmarked — no standard test for anatomical accuracy existed — but it was widely treated as evidence that the models’ persistent tells were closing rather than permanent.

The release was not without trade-offs: V5 ran more slowly than V4 on the standard four-image grid, and Midjourney’s usual “relax mode” for slower, cheaper generation was not initially available on the new version. Midjourney did not disclose architecture, training data or compute for V5, consistent with its practice on prior versions.

V5 arrived amid a wider wave of image-generator progress that spring, alongside OpenAI’s DALL-E updates and open releases built on Stable Diffusion, and it reinforced Midjourney’s reputation — despite running only as a Discord bot rather than a standalone application — as the generator most often singled out for image quality.