In 2023 I wrote Stable Diffusion for Marketers, one of the first books on Stable Diffusion for marketing. Writing about a fast-moving tool is an excellent way to practise humility in public. Plenty of the advice that circulated in that period now reads like instructions for setting the clock on a video recorder.
I was an early user of Midjourney, Runway, Stable Diffusion, ComfyUI, SwarmUI and Fooocus, so I watched a lot of that advice arrive and expire. Looking back, the habits from those years sort into two piles: the ones that died and the ones that held.
Prompt tricks aged worst.
Early image models rewarded incantations. Long strings of quality keywords, artist names stacked on top of each other, weighted brackets, negative prompts the length of a shopping list. People traded them like family recipes. Some of them helped, on that model version, for a while.
Then the models got much better at reading plain language, and most of the tricks stopped mattering. Today a clear description of the subject, the light and the framing does more than any magic word list. If your process depends on a phrase you found in a forum thread, it will break with the next update.
Settings lore went the same way.
There was a whole folklore around samplers, step counts, guidance values and seeds. Some of it was real for a given model. Much of it was confirmation bias with a spreadsheet. Interfaces have since hidden most of those dials or picked sensible defaults, and node-based tools like ComfyUI turned the useful parts into workflows you can save, share and rebuild.
The lesson is that any knowledge tied to one tool’s settings has a short shelf life. Learn it when you need it and expect to throw it away. Building an identity around knowing which button to press is a risky career plan.
References, iteration and art direction held up.
The parts that aged well were never about the software. A good reference board still decides most of the outcome. Iteration still matters: generate, pick, refine, repeat, with a human choosing at every step. Art direction matters more than ever, because a model will happily give you something attractive and wrong.
A reference board also travels between tools. When a new model arrives, the board still describes what the brand wants, and the team can test the new tool against it straight away. That portability is worth more than any setting I ever memorised.
Checking held up best of all. Hands, labels, on-screen text, faces that drift between frames, a product with the wrong cap. The tools got better and the mistakes got subtler, which makes them more dangerous, because a subtle mistake survives more meetings. That is why quality control sits at the centre of how I train teams, and why I wrote that quality control is the product.
Here is what I would tell a marketer starting now.
This year I spoke on a panel and ran a workshop at Doers Summit in Limassol, and teaching always forces the question of what still deserves teaching. My short version:
- Learn to describe images in plain, precise language: subject, light, lens, mood.
- Build a reference board before you open any tool.
- Treat every tool as temporary and every workflow as something you will rebuild.
- Keep a list of the mistakes that reach review, and add to it every time one gets through.
- Spend more time choosing outputs than generating them.
The tricks were fun. They made early adopters feel like wizards, and I enjoyed being one. The work that pays now is less magical and far more useful: knowing what good looks like, and refusing to ship anything below it. If that sounds dull, good. Dull is what clients pay for twice.