Short answer
Midjourney V4 is an all-new image model that Midjourney released as an alpha test on 5 November 2022, built on a fresh code base and trained on the company's own AI supercluster. According to founder David Holz, it knows far more about creatures and places, handles small details and crowded scenes better, and follows longer prompts, though for now it makes only square images.
Follow-up question How should you change your prompts for V4?
Midjourney users have a new model to play with. Yesterday, 5 November 2022, the independent research lab released the first alpha version of V4, its fourth image model, and the company promises more detailed, more realistic and more accurate pictures than before.
An alpha is an early test release: the core is in place, features are missing and things can still change. V4 sits alongside the existing models while testing goes on.
What Midjourney announced
Founder and CEO David Holz introduced V4 in an announcement to Midjourney's Discord community. His main points:
- V4 is not an update of V3: Holz describes a new code base and a new AI architecture written from scratch.
- It is also the company's first model to learn on Midjourney's own AI supercluster, a large computing system the company has been building for about nine months.
- It has much more knowledge, for example about creatures and places.
- It gets small details right more often.
- It can handle more complex prompts, and scenes with several objects or characters.
Holz called V4 a first step rather than the final word. He also said that because the images are now much more realistic, the community guidelines will be enforced far more strictly: people who deliberately try to slip around the list of banned words risk being banned sooner.
How Midjourney works, and where V4 fits
Midjourney runs inside Discord, a chat app. You type the /imagine command followed by a description, the bot creates a grid of four images, and you can upscale or make variations of the ones you like. Extra settings go at the end of the prompt as parameters that start with two dashes. For instance, --ar 3:2 sets a wider aspect ratio and --no grass asks the model to leave grass out.
Until yesterday, Midjourney's documentation listed V3 as the default model, in use since 25 July 2022, with V2 still available for those who preferred the original style. The docs show that a model is chosen with a version parameter or with a button in the /settings menu. Here is how the generations compare, based on Holz's announcement and the earlier documentation:
| V3 (default since July) | V4 alpha (5 November) | |
|---|---|---|
| Foundation | Midjourney's earlier code | New code base and architecture |
| Training hardware | Not stated | Midjourney's own AI supercluster |
| Knowledge of subjects | More limited | Much broader (creatures, places) |
| Small details and busy scenes | Less reliable | Handled better |
| Image shapes | Any aspect ratio with --ar | Square only, for now |
One way to picture the change: if V3 is a talented illustrator with a limited reference library, V4 is meant to be one who has studied far more of the world and pays closer attention to every part of the brief.
Current limits
Because this is an alpha, there are clear gaps. The biggest is that V4 currently creates only square images. Midjourney's to-do list includes higher resolution and image quality, more control over upscaling, sharper images, and fixes for stray text-like marks and awkwardly cropped subjects.
Holz also warned that V4 needs a different way of writing prompts. Habits that worked well with V3 may not carry over.
Try it: If you have Midjourney access, take a prompt you used with V3 and run it unchanged with the V4 alpha. Then write a second, plainer version that describes the scene as clearly as you can. Put the three grids side by side and note which details each version got right.
Why it matters
V4 arrives in a busy year for image generators. Midjourney itself only opened its public beta on 12 July, and in the months since, competing tools have kept raising expectations for what a few words of text can produce.
The bigger story is realism. A model that can produce convincing, detailed scenes is more useful for illustration, concept art and design, but it also makes misleading images easier to create, which is why Midjourney is tightening enforcement at the same moment it raises quality. And by training on its own computing cluster with a new architecture, the small lab shows it is investing in building its own models from the ground up.
How should you change your prompts for V4?
Holz's main advice is that V4 calls for a new style of prompting, and early users are still working out what that means. Because the model understands more and handles complex scenes better, a reasonable starting point is to describe what you want plainly and completely: the subject, what it is doing, the setting, the lighting and the style. Long strings of style keywords added to coax V3 into looking good may now be unnecessary, or may push the image somewhere you did not intend.
Here is a prompt written that way:
/imagine prompt: an old fisherman mending a red net on a wooden dock at sunrise, gulls overhead, soft mist on the water, detailed photographSince V4 makes only square images for now, leave out --ar and plan your compositions for a square frame. Use the remix feature, which already works with V4, to adjust a promising image instead of starting from scratch. Most important, change one thing at a time between attempts, so you can see what each word actually does. Because this is an alpha with a long to-do list, prompts that work today may behave differently as the model is improved.
Sources
- An illustrated Guide to /imagine parameters (archived 3 November 2022) — Midjourney, November 3, 2022
- Midjourney V4 can generate more concrete and realistic AI images — The Decoder (Jonathan Kemper), November 7, 2022
- Midjourney — Wikipedia, July 12, 2022