A real account of how I kept generating the wrong mug, and what finally fixed it.
Yeah, It Was a Mug 🫖
My product photo problem was embarrassingly specific: a ceramic mug, matte white, sitting on my kitchen counter in front of a drying rack, a half-visible sponge, and what I can only describe as "ambient dish guilt."
Not exactly the clean lifestyle shot I needed for a product listing.
The mug itself was fine. Nice glaze. Good proportions. The kind of thing that photographs well if you ever get around to clearing the counter first. I hadn't.
So instead of doing the reasonable thing — which apparently is to move the sponge — I decided to let AI handle it.
The Prompt Spiral 🌀
I opened a text-to-image tool and described my mug.
"A white ceramic mug on a clean wooden table, warm morning light, minimalist, product photography."
What I got back was: a mug. Technically. A different mug. Different proportions, different handle shape, a slight sheen on it that looked more "stock photo prop" than "thing someone actually made."
Fine. I got more specific.
"Matte white ceramic mug, slightly irregular rim, handmade look, on a pale oak surface, soft diffused lighting."
Different mug again. Still wrong.
I tried adding brand descriptors. I tried "artisan." I tried "Scandinavian minimalist" and then "Japanese wabi-sabi" and at some point "not the sponge" which is not a real prompt and produced predictably chaotic results.
The problem wasn't my prompting. The problem was that I was describing a mug to a model that had never seen my mug. Every generation started from zero. Every prompt was me trying to reconstruct, in words, an object that was sitting twelve inches from my keyboard.
At some point I counted: fifteen generations. Maybe more. All of them fine. None of them mine.
Every One Was a Different Mug 🪤
It took me longer than I'd like to admit to name the actual issue.
Text-to-image tools are great at producing images from concepts. They are not great at producing your specific thing from a description, because your thing has specifics that language flattens. The curve of the handle. The weight of the glaze. The way the rim sits.
I was trying to re-create a visual from words, and that process has one unavoidable failure mode: it generates the average of all mugs, not this mug.
What I actually needed wasn't a new image. I needed to keep the image I had and change everything around it — the mood, the lighting, the scene — while keeping the mug exactly as it was.
The Actual Tool for This Job 💡
Image-to-image AI works differently. You upload your photo. The model uses it as a starting point — the subject, the composition, the specific shape of the thing you actually own. Then your prompt tells it what to change.
Not "generate a mug." More like: "same mug, warm café scene, golden hour light, lifestyle photography."
That's a fundamentally different instruction. The model isn't guessing what a mug looks like. It already knows, because you showed it.
I tried imgtoimgai.io, uploaded the original photo, and wrote: "cozy café table, warm morning light, lifestyle product shot, keep the mug."
The mug it returned was my mug. Same handle. Same rim. Same glaze. But now it was sitting on a worn wooden café table with soft golden light coming through a blurred window behind it. The drying rack was a distant memory.
I did not move the sponge. I did not do another fifteen generations.
Three Things It Actually Works For 🎯
Once I understood what the tool was doing, I used it for a few other things that had been sitting in my "too annoying to fix" pile.
Product shots with wrong vibes. You have a real photo of the thing — good. But it's sitting on a kitchen counter under fluorescent light and looks like a police evidence photo. Upload it, describe the scene and mood you actually want — a café table, a studio shelf, a sunset patio — and the model re-renders the whole atmosphere while keeping your specific product in frame.
Portraits in the wrong setting. I had a decent headshot taken in a parking garage. The face was fine. Everything else looked like a parking garage. Asked for a warm indoor portrait with soft natural light. Two minutes. Good enough for a speaker bio, which was the whole bar.
Sketches turned into reference images. I'd roughed out a layout for a product label — more of a diagram than a drawing. Uploaded it, described the finished style I was going for, and got something close enough to drop into a brief and show a designer. Not the final design. A useful, specific starting point instead of a vague description.
None of these outputs were perfect. That is not the point. The point is that the object in the output was recognizably the object in the input, which is the one thing that text-to-image genuinely cannot give you.
"Good Enough" Is the Actual Bar ðŸ§
My mug listing went up with a café-scene lifestyle photo. Nobody has mentioned the kitchen counter. A few people have mentioned the mug.
I probably spent forty-five minutes on the whole thing, including the fifteen generations I did before switching tools. If I'd started with image-to-image, it would have been ten.
I am not a designer. I do not know what most of the settings do. My version of "product photography" is "doesn't look like I took this next to a sink."
But that bar, it turns out, is meetable.
Try it at imgtoimgai.io — upload your photo, describe the change, get your thing back.
Top comments (0)