Midjourney and generated imagery
Generated images are cheap and fast, and they are right for some jobs and wrong for others. The decision worth making is not whether to use them. It is which surfaces they are allowed on, who signs off, and what you can still prove about an image a year later.
01Capabilities
Four places generated imagery earns its keep
Midjourney is an image generator, reached through its web app and through Discord, and it is unusually good when a person is steering it. That shapes where it fits: exploration and illustration with a human in the loop, never unattended production. These four are the cases we use it for inside product design.
- Direction
Three credible visual directions in an afternoon
So the conversation about tone happens against something rather than against adjectives. It is an input to the design work and not the output, and we label it that way, so nobody mistakes an exploration frame for a committed direction.
- Editorial
Imagery that sets a tone rather than states a fact
Article headers, section illustrations and deck imagery, where the picture carries mood and nothing else. The rule we apply is short: if a reader could reasonably take the image as evidence of something, it does not get generated.
- Prototypes
Placeholders that stop distracting the reviewers
A prototype needs imagery plausible enough that people judge the interface instead of the grey boxes. Generated assets are a better answer than borrowed stock or, as still happens, a competitor's photographs pasted into a demo.
- Pipeline
The unglamorous half: naming, storage, approval
At any volume the engineering is the boring part: naming, storage and delivery, a record of what produced each asset, a review queue with a named approver, and alt text written by a person. Producing images inside a running product is a separate AI integration decision.
02Fit
Surface by surface, where we allow it and where we do not
The question is never whether the tool is impressive. It is whether a generated image is acceptable on a specific surface, for a specific audience, with a specific consequence if it turns out to be wrong. We answer that surface by surface, in writing, before anybody opens the tool. Five of the six rows below send you somewhere other than a Midjourney project with us, and one of those names no unicrew page at all.
| Your situation | What we recommend |
|---|---|
| Exploring visual direction early, where the quality of the aesthetic is what matters | Use MidjourneyIts strongest case, and the one we reach for most. A designer is steering, iterating and rejecting, which is the condition it is good under. |
| Images have to be produced inside your product, or in a batch job | Different pageUse a model with a documented, supported API, such as OpenAI or Gemini, and keep a person approving anything a customer will see. |
| You sell the thing in the picture: a product, a property, a room | Photograph itA generated image of real inventory is a returns and trust problem in e-commerce, and a disclosure problem in property, long before it is a legal one. |
| One consistent character or spokesperson across a whole campaign | Commission itOutput drifts between runs, and the effort spent forcing consistency usually costs more than the illustration or the shoot would have. |
| Anything showing patients, clinical detail, financial outcomes or people who look like your customers | We keep it outIn healthcare and fintech the downside is lopsided. Nobody praises an illustration, and one implausible detail costs credibility you spent years earning. |
| Icons, diagrams and interface illustration that have to match your product | Different toolVector work in Figma, against your design system. Raster generation cannot hit a token, a grid or a stroke weight. |
Scope
What we own here is the policy and the trail, not the prompt craft. A written allow and deny list per surface, agreed with whoever owns your brand and whoever owns your risk. A record on every asset, so that a year later you can still tell what was generated. Alt text written by a person, because a generator does not know why the image is there. A named human approving anything that reaches the public. unicrew has been building custom software since 2012, with 100+ senior in-house engineers across six countries. Design and delivery work runs under an ISO 27001:2022 certified information-security management system. Where the answer is that generated imagery is not acceptable, we say so and budget for the commission.
03Delivery
How this runs on a live project
Four steps, and the first two happen before anybody opens the tool. Skipping them is how a team ends up arguing about an image that has already been published.
- Agree where generation is allowed, in writingA surface-by-surface list: allowed on article headers and internal decks, for example, and banned on product photography, case studies, anything clinical, and anything a reader could take as evidence. Agreed by brand and by risk together, so it is not relitigated per ticket.
- Settle the terms before anyone generates anythingRead the current plan terms rather than a summary, because they set conditions by tier and by company size and they get revised. Then decide what visibility of your prompts and outputs you can accept on that plan. Your own counsel confirms this, not us.
- Build the review gate and the asset trailNothing reaches a public surface without a named approver. Every asset records what produced it and when, lives in your asset store rather than in somebody's downloads folder, and ships with alt text a person wrote for the place it is used.
- Keep the commissioned option fundedBudget for real photography or illustration on the surfaces where consistency and truth matter. Teams that skip this step end up either shipping something they are unhappy with, or paying for the shoot late and in a hurry.
04Stack
What produces the rest of the imagery
The tools that make everything a generator should not, and the two providers with an API you can build on. Each has its own page.
05Questions
What teams ask before letting generated images ship
The questions a brand owner and a risk owner both ask, answered together.
You can bring our designers onto your team, and we will not hand over people and walk away from what gets published. Work goes through the same gate as our own: a design lead outside the delivery team reviews what ships, generated assets carry their record, and alt text is written by a person. Capacity is managed teams; an outcome is product design.
The vendor's terms decide what you may do with the output. Those terms set conditions by plan tier and by company revenue, and they get revised, so read the current version and have your own counsel confirm it. Separate two things: permission to use an asset is not ownership of copyright in it. In the United States the Copyright Office has said work produced without human authorship is not registrable. If exclusivity matters, commission it.
It comes down to whether a person is in the loop. Midjourney is strong when a designer is steering, iterating and rejecting, which is what exploration and editorial illustration look like. When images have to be produced inside your product, on a schedule or at volume, you need a documented, supported API and terms that cover your use, so OpenAI or Gemini instead.
Reference images, tightly controlled prompts kept in one place and a fixed set of style parameters get you a long way, and not all the way. Output still drifts between runs, and anyone claiming to have solved brand consistency in a generator is overselling it. So we solve it the other way round: cap which surfaces use generated imagery and commission the rest.
A scoping call. We want to know which surfaces the imagery is for, who approves what gets published, and whether anything you are considering touches a regulated or evidential claim, because that usually redraws the list on its own. What comes back is a written allow and deny list with the reasoning. Most engagements start within two to four weeks.
Three shapes, and which fits depends on how settled the scope is. Time and materials is billed hourly and quoted per project, which suits work still moving. Fixed price is outcome based, offered once the first read is done, because a fixed number before anybody has seen the surfaces is a guess with a contract around it. Team extension is billed monthly per engineer.
Working out where generated imagery is acceptable?
Bring us the surfaces and the audience. You get a straight allow and deny list, including the places where we would tell you to pay a photographer.