Midjourney V7 Review: Beautiful, Opinionated, Still Not Free
V7 adds personalisation and Omni Reference, closing the two gaps that used to send people elsewhere. It is still the wrong tool if you need the image to be accurate rather than striking.
There is a particular moment that happens to almost everyone who tries Midjourney for the first time. You type something vague, you get back an image far better than what you described, and you think: this is going to be easy. Then you try to make a specific image — a particular character, in a particular pose, in a particular style — and you spend forty minutes discovering that the model has opinions and they are not yours.
Version 7 is mostly about that second experience. Midjourney has been the best-looking image generator for years without being the most controllable one, and the V7 release is aimed squarely at the gap between “beautiful” and “what I actually asked for”. This review covers what changed, what it costs, and where it still loses to its rivals.
Why AI image generation became a real workflow
The interesting shift was not artistic, it was economic. Stock photography was always a compromise — you searched for the image in your head, failed to find it, and settled for something adjacent. Commissioning illustration solved that but cost hundreds of pounds and took a week.
Generation collapsed the gap. The image in your head is now roughly twenty minutes of iteration away, at a marginal cost close to zero. For blog headers, concept art, mood boards, pitch decks, book covers and product mock-ups, that has genuinely changed how the work gets done.
It has not replaced photography or illustration at the top end, and the reasons are instructive: generated images are unreliable about specifics. Real products, real people, real logos, text, hands doing precise things, brand consistency across a campaign. Anywhere accuracy matters more than atmosphere, you are still better off with a camera or a person.
The types of image model, and why Midjourney sits where it does
Aesthetic-first models — Midjourney being the definitive example — are tuned to produce beautiful output from lazy prompts. They make opinionated choices about lighting, composition and colour. This is why a two-word prompt looks like a film still.
Instruction-following models — the image generation built into the large chat assistants, and several open models — are tuned to do what you said. Output is often plainer, but if you ask for a red mug on the left, you get a red mug on the left.
Open, self-hosted models trade convenience for total control: custom training, no content filters you did not choose, no subscription, but a real technical setup cost.
Which explains the most common complaint. People who find Midjourney frustrating are usually trying to use an aesthetic-first model as an instruction-following one. If your prompts read like a specification, you will fight it. If they read like a description of a mood, it will astonish you.
What V7 actually changed
Personalisation profiles. The model learns your aesthetic preferences from your own ratings and applies them by default. This is the feature that most changes daily use: after a few hundred ratings, your default output starts arriving closer to what you would have picked anyway.
Omni Reference. You can hand the model an image and have it carry both subject and style into new generations. This is the direct answer to the biggest historic weakness — keeping a character or object consistent across a set of images. It is not perfect, but it is the difference between “possible with effort” and “not really possible”.
Draft Mode. A much faster, lower-fidelity generation pass for exploring ideas before committing. Since most of the work is discarded iterations, speeding up the discarding is worth more than it sounds.
Wider aspect ratios, up to 4:1, which matters for banners, panoramas and cinematic framing that previously needed outpainting or a crop.
What it costs
Midjourney is subscription-only. There is no permanent free tier, which is unusual in this category and is the first objection most people raise. Prices below are the published plans as of August 2026, with a discount of around 20% for paying annually.
| Plan | Monthly | Roughly who it suits |
|---|---|---|
| Basic | $10 | Trying it seriously, occasional use. Fast generation time only, so you will feel the limit. |
| Standard | $30 | The real starting point. Adds unlimited generation in the slower relax queue, which changes everything about how you iterate. |
| Pro | $60 | Heavy daily use, plus private generation for client work. |
| Mega | $120 | Studios and high-volume production. |
The jump worth understanding is Basic to Standard. On Basic you are metered on fast generation hours, and because good results come from iteration, metering iteration is the thing most likely to make you feel the tool is not working. Standard’s relax queue is slower per image but effectively unlimited, and that removes the anxiety that makes people prompt badly.
If you are deciding: a month of Standard tells you more than three months of Basic. The tool rewards volume, and Basic is the plan most likely to leave you thinking you cannot use it.
Where it still loses
Text inside images. Better than it was, still the weakest area. If your image needs a readable word in it, generate the image and add the text yourself.
Literal spatial instructions. “Three objects, the blue one on the left” is the sort of prompt this model reads as a suggestion.
Brand consistency. Omni Reference helps a great deal, but producing forty images that look like the same campaign is still work rather than a setting.
No free trial. You cannot evaluate it without paying, which is a genuine barrier and pushes a lot of casual users to rivals permanently.
What it does well
- Still the best-looking output in the category by a clear margin
- Personalisation makes defaults progressively more useful
- Omni Reference finally addresses subject consistency
- Draft Mode makes exploration genuinely fast
- Web app removed the Discord barrier that put many people off
What to think about first
- No free tier at all — you pay to evaluate
- Basic plan’s metering makes the tool feel worse than it is
- Text rendering still unreliable
- Resists precise spatial instructions
- Commercial and licensing terms need reading if you have clients
How it compares
| Tool | Strongest at | Weakest at | Choose it if |
|---|---|---|---|
| Midjourney V7 | Visual quality, style, atmosphere | Literal accuracy, text, no free tier | The image needs to look striking |
| Chat-assistant generators | Following instructions, text in images, iterating conversationally | Distinctive look | The image needs to be correct |
| Adobe Firefly | Commercial safety, Creative Cloud integration | Rawer output quality | Licensing clarity is non-negotiable |
| Open models (Stable Diffusion family) | Control, custom training, no fees | Setup and maintenance | You want to own the pipeline |
Who should pay for it
You should, if you produce visual content regularly and the look of it is part of the value — content marketing, book covers, album art, concept work, pitch decks, editorial illustration. Also if you are a designer using it for ideation rather than final assets, which is where it is genuinely at its best.
You should not, if you need images that are factually accurate about real things, if you need readable text in the image, or if you generate three pictures a year. In all three cases something else serves you better.
Working with images?
Generated images usually need converting, compressing or resizing before they go anywhere. Our free browser-based image tools do all of that without uploading anything to a server.
Frequently asked questions
Is there a free way to try Midjourney?
Do I still need Discord?
Can I use the images commercially?
Which plan should I start on?
Why does it ignore parts of my prompt?
Can it keep the same character across several images?
The verdict
Still the one to beat on pure image quality, and V7’s personalisation and Omni Reference close the two gaps that used to send people elsewhere. Start on Standard, not Basic. Use it for anything where atmosphere matters, and use something else for anything where accuracy does — that division has not changed and probably will not.
Sources and method
Plan names, prices and feature descriptions reflect Midjourney’s published information as of August 2026 and change without much notice. Several third-party reviews quote precise percentage improvements for V7 over V6; we have not repeated those, because they do not trace back to any published benchmark. Descriptions of model behaviour here reflect documented features and consistent reporting across independent reviews. We have not run controlled comparison tests ourselves and do not claim to have.