Midjourney

Midjourney is the image model that decided aesthetics were a feature rather than an emergent property, and every version since has been tuned toward a house style rather than toward neutrality.

The opinionated model

Ask three image generators for “a woman reading in a cafe” and Midjourney’s will be the one that looks composed. Depth of field where a photographer would put it, light that falls from somewhere specific, a colour palette that holds together. The model has taste, and that taste is consistent.

For anyone producing concept art, mood boards, editorial illustration or marketing imagery, this is exactly the right trade. A default that already looks intentional means fewer iterations to something usable, and the floor is much higher than models that require careful prompting to avoid looking generic.

The cost of the same decision is that the style is difficult to escape. Midjourney images are recognisably Midjourney images, and after two years of saturation that is a liability for brands trying to look distinct rather than current.

Control has improved substantially

The early criticism — beautiful and uncontrollable — no longer holds. Style references pin an aesthetic across a series. Character references keep a figure consistent between images, which is what made narrative and campaign work viable rather than a lottery. Region-based editing lets you fix the one wrong element instead of rerolling the entire image and losing everything that worked.

Parameters for stylisation strength, aspect ratio, chaos and weirdness give real dials, and learning them is the difference between casual and professional use of the tool.

Where it still loses

Text rendering remains weak. If your image needs legible words in it, this is the wrong model and no amount of prompting fixes it — composite the type afterwards.

Photorealism is deliberately not the goal. Midjourney’s realism is cinematic rather than documentary, and it reads as a photograph shot by someone with a very expensive camera and a strong point of view. For product photography or anything requiring the image to be unremarkable, competitors do better.

Precise compositional control — this object here, that one there, at this size — is still awkward compared to tools built around layout. Midjourney interprets; it does not take dictation.

And the commercial-use question needs answering by your own legal team rather than by a review. Training data provenance for image models remains contested, the licensing terms attach to your subscription tier, and “we generated it with Midjourney” is not by itself a clearance.

Pricing

Subscription only, in several tiers that scale by generation volume and by whether your work is private. There is no free tier, which is unusual in the category and is a deliberate filter.

For a professional producing images regularly the lowest paid tier is trivially justified. For occasional use it is harder, and the absence of a free path means you cannot properly evaluate the aesthetic fit before paying — run one month at the entry tier and decide from real output.

Who should use something else

Teams needing text in images, precise layout control, or integration with an existing design workflow should look at Adobe Firefly, which trades aesthetic ambition for commercial safety and a place in Creative Cloud.

Teams that need to self-host, fine-tune on their own material, or generate at volume through an API should use FLUX or another open-weights family.

Teams that want image generation bundled with everything else rather than as a separate subscription will find ChatGPT adequate for most non-specialist work.

Our full comparison of the field is at best AI image generators.