Marigold was a one-day concept for a design challenge. I set the direction in real photography before generating anything. That order matters more than people think. If you generate first, you're negotiating with whatever the model offers. If you shoot or choose real references first, you've set a standard the rest of the work has to meet. The tool didn't set the taste. It matched a bar that already existed.

Great work!, love the colors

When was the first time to use a generative AI for visuals? Some OG stuff, I will go first: April 2023, using Midjourney in Discord /imagine... These felt like magic back then. You really had to learn how to write good prompts. It was hard to get consistent results with characters and scenes. Now all AIs handle everything for the user. Visuals look more realistic than photographs, yet I am not impressed. Somewhat after so many years, the tech became smooth and perfect, but flat and boring. Magic is gone for me, now it is another wide-spread tool everyone uses. (some 2023 generations below)

Late 2022 for me, DALL·E 2, mostly moodboard filler that never went near a client. What you're calling flat might be an authorship problem rather than a quality one — back then the model failed in odd ways and you had to steer it hard, so whatever survived carried your...