Two Tuesdays ago I needed a hero illustration for a product launch page. The brief was specific. A warm, slightly analog desk scene. Laptop, coffee cup, plant. Three brand colors. The new piece had to sit next to six existing spot illustrations on the marketing site without anyone noticing it was the new one. Time budget: one afternoon, including revisions.
I opened Midjourney. Forty minutes later I had the asset that shipped. Not a reference, not a mood board. The actual hero, in the actual layout, on the actual page that went live that week.
This would not have worked in 2024. It barely worked in summer 2025. Sometime in early fall, it started working.
I want to talk about what changed, what it unlocks for working designers, and where it still falls down. I’m not a generative-models researcher and I’m not going to pretend I am. I’m someone who opens a design tool every day for a paying job, and I’m telling you what I noticed.
What actually got better
So what actually changed between summer and fall? If you track image-generation tools at all, the jump from Midjourney v6 to Midjourney v7 was the first one where the output cleared the bar for product work in a way I’d actually ship. Not “cherry-picked demo” reliable. “Ship-an-asset-this-week” reliable.
Google’s Imagen 3 and OpenAI’s image generation in GPT-4o made similar jumps in the same window. Flux from Black Forest Labs hit a different quality axis and became the go-to for designers who needed real control over composition. The whole category leveled up at once, and the leveling was actually enough to push the tools over the line for product work.
That line has three parts.
First, typography that’s usable. The models were historically a disaster at rendering text. Wrong spelling, weird kerning, none of it surviving a localization pass. That’s still partially true. But “button labels that say button” and “headings that spell the right word” have gotten close enough to use in comps. You still shouldn’t render production text inside a generated image. You can now use generated images with text in exploration without burning twenty cycles on the spelling, which is, honestly, a meaningful change.
Second, consistency across a set. The old failure mode was that you’d generate a beautiful hero and then spend two weeks failing to make three companion illustrations look like they belonged to it. Midjourney’s style reference codes and character reference have gotten good enough that you can establish a brand look and generate within it. This is the single biggest unlock for UI work. You don’t just need one image. You need a family.
Third, color control that respects brand. Early generative work was loose with color. You’d prompt “teal and orange” and get something that trended that way but had seventeen other colors hiding in it. The palette parameters in v7 and the equivalents in Flux and Ideogram closed enough of that gap that your brand team won’t reject the output on sight. Not perfectly. Enough.
How my desk looks different
Here is what asset production looked like in 2023.
Open Google Images. Open Dribbble. Open Pinterest. Open Are.na. Spend two hours hunting reference that’s close to what I want but not quite. Capture the good pieces into a Figma mood board. Start sketching. Realize the reference I collected is stylistically incoherent. Restart the mood board. Three hours in, draft an actual first version of the asset. Iterate. Ship after another four hours.
Here is what it looks like now.
Open Midjourney. Write a long, specific prompt. Generate sixteen variations. Notice five of them are interesting. Pick the two closest to the brief. Lock the look with a style reference. Generate twelve more variations with the look locked. Pick the best one. Open it in Figma, do a color and composition pass, drop in any real text that needs to be real, ship.
Roughly 60 percent less time. Output usually better, because the constraint-rich prompt forces me to name what I want instead of finding it through trial.
The model doesn’t replace the designer. It multiplies whatever the designer puts into it. That’s a bigger lever for taste than any tool we’ve had.
The interesting thing isn’t really the time saved. It’s what moves in the workflow. The reference-hunting phase used to be where I did my visual thinking. I’d find a thing that worked and it would teach me what I actually wanted. That phase is gone. It’s been replaced by a prompt-writing phase where I have to know what I want upfront, which is a different muscle. Designers who can name what they want in words are faster. Designers who need to see something to know what they want are slower. That’s a real shift in which design skills compound, and I think it’s one of the bigger structural changes that happened to the discipline this year.
Where it still falls down
Licensing is unclear and it bothers me. Midjourney’s terms say you own commercial rights on a paid plan, but the underlying training data is in active litigation (the Getty case is the obvious one), and I don’t know how that shakes out. I’m using the tool for commercial work anyway because that’s the category norm, but I’m not comfortable about it. If you work at a company with a risk-averse legal team, they should be paying attention. If you work in legal at one of the model labs and I’m getting this story wrong, please let me know.
Model-to-model consistency is still weak. If you produce a set of assets on Midjourney v7 and then v8 ships with a slightly different aesthetic, you may have to regenerate the whole set to keep them feeling related. That’s a real operational cost the industry hasn’t fully priced in. The playbook for keeping a visual library coherent across model generations doesn’t exist yet. It’s going to.
“Everything looks like Midjourney” is a problem. The house style of these models is strong, and it leaks into everything that comes out of them. A skilled prompter can push past it. A casual user cannot. The market is flooding with content that has a recognizable generative aesthetic, and if you care about brand differentiation, the move is to use these tools as a starting point and finish by hand. The designers who finish by hand will have more distinctive work. The designers who don’t will produce the visual equivalent of 2014’s Bootstrap websites. Functional. Indistinguishable.
A practical tip
For designers who want a brand-consistent generation set, here’s what works for me.
Produce three reference images by hand, in the exact style you want. Not by generating and tweaking. Actually draw or collage them. These are your anchors. Feed them to Midjourney as a style reference triplet. Then generate everything downstream with that reference locked in.
This inverts the usual pipeline. Instead of using the model to produce and iterating visually, you use your own hand-craft to set the style and use the model to multiply it. The output is more distinctive because the seed was human. The cost is that you still have to be able to produce three reference images by hand, which is a real skill. The upside is that those three references are reusable across a year of asset production, and they propagate your team’s actual visual identity rather than Midjourney’s.
Honestly.
That’s the whole shift, in one tip. The model is a multiplier. The taste is yours.
This is the last retrospective essay I’m writing to fill the gap between when things actually happened and when I started diffuse.design. Starting next week, we’re live in the present. I’ve been thinking a lot about what 2026 is going to mean for the profession, and I’ll have more to say about that in the new year.
See you in two weeks.
Jameson