What is new in Midjourney V5, and how do you use it?

Yesterday, 15 March 2023, Midjourney released version 5 of its image generator. Within hours, people were posting pictures that look like professional photographs: portraits with visible skin pores, food shots with realistic reflections, scenes lit like a film set. For many, the most talked-about detail is smaller. Midjourney's people now usually have hands with five fingers.

V5 is an alpha, an early test version that may still change. It is open to paying subscribers, and the older V4 stays the default unless you ask for the new one. Here is what changed, how it works, and how to try it.

What happened, and who makes Midjourney

Midjourney is a small independent research lab in San Francisco, led by David Holz, a co-founder of the hand-tracking company Leap Motion. Unlike most image tools, it has no app of its own for making pictures. You use it inside Discord, a chat platform, by typing commands to a Midjourney bot, which replies with four images.

The lab has moved quickly. Its own documentation lists version 1 as the default from February 2022, followed by versions 2 and 3 that year; Ars Technica dates Midjourney's first arrival to March 2022. Version 4, built on a new codebase and trained on Midjourney's own cluster of computers, became the default in November. V5 comes about four months later.

What is new in V5

Midjourney's documentation calls V5 its most advanced model so far. Combining that page with the release notes as reported by Ars Technica, the main changes are:

Area	V5 alpha, compared with V4
Realism	Much closer to photography: skin, faces, lighting, reflections and shadows
Hands	Usually drawn with five fingers
Prompts	Better at reading natural sentences, and more sensitive to wording
Style range	A much wider range of styles
Resolution	About twice the resolution
Stray text	Less unwanted lettering inside images
Patterns	Repeating tiles with the --tile parameter

Early users are describing the jump in vivid terms. The designer Julie Wieland told Ars Technica that hands are now "correct most of the time," with five fingers rather than seven to ten. Some fans joke that the results are almost too perfect, which removes part of the fun of rolling the dice until a good image appears.

How it works, and how to use it

Like DALL-E 2 and Stable Diffusion, Midjourney is a text-to-image model: a neural network trained on very large numbers of pictures paired with descriptions, so it learns which visual patterns go with which words. When you write a prompt (your description), the model builds a new image that matches those patterns. Midjourney has not published the inner design of its models, so how V5 differs internally from V4 is not public; what users can see is the output.

Choosing a version is simple. You either add a parameter, a short setting typed after the prompt, or change your default with the /settings command. Here is a worked example you can run as a subscriber:

In a Midjourney channel or a direct message with the bot, type /imagine and press space.
Enter your description and add the version parameter at the end:

Prompt

a weathered fisherman mending a net at dawn, soft window light, close-up of his hands, 35mm photograph --v 5

Wait for the bot to return its grid of four images.
Run the same prompt again with --v 4 instead, and compare the hands, the skin and the light.
Try writing the description as a full sentence rather than a list of keywords. V5 is designed to respond better to plain language.

To make V5 your default, type /settings and choose "MJ Version 5." Earlier versions remain available through --v 1 to --v 4.

Try it: Pick one scene and write it two ways: a list of keywords, and a single natural sentence. Generate both with --v 5 and note which version follows your intent more closely, and which details it still gets wrong.

What should you watch for now that AI images look this real?

The first thing to watch is your own trust in photographs. One of the simple tells of AI images over the past year has been mangled hands. With V5, that clue is fading. A widely shared post this week warns people to be extra critical of political images online, especially ones designed to provoke anger. Good habits help: check who published a picture, look for the same event from other sources, and be wary of dramatic images with no clear origin.

The second is ownership. Midjourney has recently been at the center of a copyright dispute over a comic book made with earlier versions, and the question of who owns AI-made images is still being worked out. If you plan to use V5 images commercially, read Midjourney's terms and keep an eye on how that debate develops.

The third is the alpha label itself. Midjourney is testing V5 in public, so results, settings and behavior may change in the coming weeks. If a prompt gives strange results, try it again in V4 and compare. Noting what each version does well is the fastest way to learn the tool.