Put AI on a meeting agenda and somebody will ask for the tool list before the coffee arrives. Which model, which settings. Fair enough. People like shopping lists. The trouble is that the list explains very little about the work that comes out the other end.
I was an early user of Stable Diffusion and Midjourney, and I have spent a lot of hours in Runway, ComfyUI, SwarmUI and Fooocus. In 2023 I wrote a book on Stable Diffusion for marketers. Plenty of people had the same tools at the same time as me. Their work looked nothing alike, and the difference had very little to do with software.
The tool is the cheapest part of the setup
Most of these tools are available to anyone with a card and a login. Your competitor has them. So does your intern. Whatever edge a tool gives you lasts until the next update, and then everyone gets the same edge. What stays scarce is the person who knows what good looks like for this brand and can get there on purpose.
Output quality depends on four things the pricing page never mentions:
- Taste. Knowing which image in a pile of decent ones is the right one, and being able to say why.
- References. Starting from real lighting, real products and real art direction, so the model has something to aim at.
- Iteration. Changing one thing at a time and keeping track of what moved the result.
- Checking. Looking at hands, labels, text, edges and continuity before the client does.
None of those four come with a subscription. They come from practice, from looking at a great deal of bad output, and from being willing to throw away a good-looking image because it is wrong for the brief.
How to judge an operator in a portfolio review
A portfolio full of finished images tells you the operator can pick a nice frame. It tells you nothing about how many attempts it took or whether they can do it again on a deadline. So ask about the process behind two or three pieces. Who wrote the brief? What references did they start from? What did the first version look like, and what changed between that and the final?
A good operator answers with specifics and shows the ugly middle stages without embarrassment. A weak one talks about the tool. If every answer starts with the name of a model, you have met a fan. Fans are lovely. You still should not hand them your launch.
Set a test brief that shows the thinking
The best test is small and slightly annoying. Give the candidate a real product with a real label, a short brief and a handful of references. Ask for two finished visuals plus a short note on what they tried, what failed and what they checked before sending. Then read the note next to the work.
If you want to see how they cope with real conditions, change one detail of the brief halfway through, the way clients do. The good ones adjust, and they tell you what the change cost.
Look for whether the label survived and whether the references shaped the result. Look hardest at whether they spotted their own mistakes. Someone who flags a warped logo in their own submission is worth more than someone who sends a flawless-looking image with the brand name misspelt. Catching errors is a skill. Admitting them is a rarer one.
Hire the hands, then pick the tools together
Once you trust the person, let them choose the tools and explain the choice. The stack will change. It has changed many times since I started, and it will change again before this post gets old. The habits that make the work good carry over from one tool to the next, and the habit people drop first is checking. I wrote about that in why every AI team needs a quality department.
So the next time someone sends you a ranking of the top AI tools, read it if you like. Then go and find out who is holding them.