It was a Thursday evening. The kind where the sky outside my window goes from orange to gray before I notice, and I'm still sitting at my desk — the one wedged between a bookshelf and a wall that I keep telling myself I'll rearrange — staring at a folder of client assets I've been reorganizing for the past hour instead of actually working.
I opened an old project folder by accident. A brand identity job from about fourteen months ago. The image assets were a mess: stock photos, half-edited PNGs, three different aspect ratios, a Figma export that nobody asked for. I remembered how long that project took. I remembered thinking, at the time, that there had to be a better way to generate visual references quickly without paying for stock licenses or waiting on a photographer.
That was before I started seriously using an AI photo generator.
So I sat there in the blue-gray light of my monitor and thought: okay, let me actually audit this. Did integrating AI image tools into my workflow make things better, or did I just add a new layer of chaos on top of the old chaos?
Honest answer: mostly the second one. But not entirely.
Failure #1: The Prompt That Sounded Specific But Wasn't
The first real project where I tried to use an art photo generator as a core workflow tool was a series of editorial illustrations for a small online magazine. The brief was clear enough — "urban loneliness, contemporary, muted palette, no faces" — and I thought that translated cleanly into a prompt.
It did not.
I spent the better part of two evenings generating variations. The problem wasn't that the outputs were bad. Some were genuinely striking. The problem was that "muted palette" means something different to a diffusion model than it means to a human art director. I got images that were technically desaturated but compositionally loud — busy backgrounds, aggressive cropping, the kind of visual tension that reads as anxiety rather than solitude.
I kept adding qualifiers. Quiet. Still. Empty streets. Soft light. Each addition helped a little and broke something else. By the time I had something the editor approved, I had a prompt that was 94 words long and had been revised eleven times. That's not a workflow improvement. That's a new kind of homework.
What I learned: Prompt length is not a proxy for prompt precision. You can write a hundred words and still be describing a feeling rather than an image. Anyway, the piece ran. The editor liked it. I was too tired to feel good about it.
Failure #2: The Consistency Problem Nobody Warned Me About
Second attempt. A different client, a product-adjacent project — they wanted a set of lifestyle images showing their physical product in various "authentic" environments. I figured this was exactly what AI image tools were built for.
The outputs were individually fine. Put them side by side and something was off. The light direction shifted between images. The grain texture changed. The color temperature drifted by maybe 200K between the "morning kitchen" shot and the "evening desk" shot, which sounds small until you see them in a layout together and your eye immediately knows something is wrong without being able to say what.
I tried using reference images to anchor the style. I tried writing style descriptors into every prompt. I tried generating everything in one session to minimize model drift. None of it fully worked.
The client didn't notice, or didn't say anything. I noticed. I spent an extra four hours in Lightroom doing manual color correction on AI-generated images, which is a sentence I did not expect to be writing in this decade.
What I learned: An AI photo generator is excellent at producing one good image. It is not, at least in my experience, a reliable system for producing twelve images that feel like they came from the same shoot. Prompt engineering gets you partway there. The rest is still manual. Anyway, the project shipped.
Failure #3: The Prompt Engineering Rabbit Hole
By the third project, I had started reading about prompt weighting — the idea that you can assign relative importance to different elements of a prompt using syntax like (keyword:1.4) or [keyword] depending on the tool. I got interested. Maybe too interested.
I built a small personal reference doc. Tested weight values. Read forum threads. Watched a forty-minute video by someone who clearly knew more than me and used terminology I had to look up. I started treating prompt engineering like a technical discipline, which it arguably is, but I had crossed the line from "learning a tool" into "the tool is now my hobby."
The project itself — a set of atmospheric reference images for a game studio's mood board — came out well. But I had spent roughly six hours on prompt research for a job that billed at three. The images were good. The economics were not.
What I learned: There's a version of prompt engineering that makes you faster, and a version that makes you feel productive while actually slowing you down. I was firmly in the second version. Knowing the difference requires a kind of self-awareness I apparently don't have at 7pm on a Wednesday.
The One Time It Actually Worked
Fourth project. A friend asked for help with a personal zine — no budget, no deadline, no brief beyond "something that feels like late autumn in a city you've already left." She sent me a voice memo describing it. Twelve seconds long.
I opened Photogenerator, typed something close to what she'd said — almost verbatim, no technical qualifiers, no weight syntax, no reference images — and generated a batch of twelve.
Three of them were exactly right. Not technically perfect. One had a slightly warped railing in the background. But the feeling was there: that specific melancholy of a place you remember more clearly than you experienced it.
My friend cried a little. (She's allowed to. It was her zine.)
I think what worked was the absence of optimization. No prompt engineering. No consistency targets. No client approval loop. Just a description of a feeling, given to a tool, without expectation.
What I learned: The workflow that works best with an AI photo generator might not be the most technically sophisticated one. Sometimes the right prompt is just an honest sentence.
Where I've Landed
I still use AI image tools. I've stopped trying to make them the backbone of a production workflow and started treating them as a fast sketching layer — something I use to externalize a visual idea quickly before deciding whether it's worth pursuing with more controlled methods.
The prompt engineering knowledge isn't wasted. But I hold it more loosely now. I use it when precision matters and ignore it when feeling matters more.
Fourteen months ago, looking at that chaotic folder of stock photos and Figma exports, I thought AI tools would simplify things. They didn't simplify anything. They added a new set of tradeoffs.
But maybe that's the only honest thing you can say about any tool worth using:
It doesn't remove the hard part. It just moves it somewhere you haven't looked yet.
Seed combination used: 2 · 4 · 1 · 5 · 2 · 8 · 5 · 9
Top comments (0)