You generate a product shot, look at it, and something feels off. You can't name it. The bottle is there, the label is readable, the background is clean. And yet it looks like a drawing of a product rather than a photo of one. Most people assume the fix is a better model or a longer prompt. Usually it's neither. The problem is that the image is missing the boring physical evidence that cameras capture whether you want it or not.
If you want the full tool first, you can open the image generator here and follow along. Below is what to change once you're in.
Flat light is the number one tell
Beginners describe objects. Photographers describe light. When a prompt says "a ceramic mug on a wooden table," the model has no reason to invent a light source, so it lights everything evenly from everywhere. Real rooms don't do that. Light comes from a window, a lamp, a doorway. It hits one side and leaves the other darker.
Instead of naming the object, name where the light is coming from and what kind of light it is. "Late afternoon light from a window on the left, soft shadows falling to the right" gives the model an actual scene to build. You'll see the shape of the object start to read properly, because a form only looks three-dimensional when part of it is in shadow.
Nothing is touching anything
Look closely at AI product images and you'll notice products hovering slightly above surfaces, or sitting on them with no shadow underneath. In real photography there is always a contact shadow, that dark smudge right where the base meets the table. It's subtle and most people never notice it consciously, but its absence is what makes an image feel pasted together.
You can ask for it directly: "a soft contact shadow where the base meets the surface." Two extra words, and the object stops floating. If the product is transparent, like glass or a bottle, add "light refracting through the glass onto the table" and the whole thing settles into the scene.
Every edge is too sharp
Generated images tend to render everything in perfect focus, from the front label to the back wall. Camera lenses don't work that way, especially up close. A macro or product shot has a shallow depth of field, meaning only the subject is sharp and the background falls off into blur.
Adding "shot at 85mm, shallow depth of field, background softly out of focus" does more for realism than any amount of detail-prompting on the product itself. It's the single strongest lever I know of for making a render feel photographed, and it costs you nothing in prompt length.
Surfaces are too clean
Real materials have imperfection. Wood has grain that isn't perfectly even. Metal picks up fine scratches and fingerprints. Fabric has fuzz. A brand-new render is often eerily pristine, which reads as computer graphics even when the viewer can't say why.
Rather than piling on adjectives, pick one or two specific imperfections that belong to the material. "Weathered oak with visible grain and a small water ring" beats "beautiful rustic wood table" by a wide margin, because the first one gives the model something concrete to draw.
The label text is the giveaway
Text is where generated images most often fall apart, and product shots are full of it. If your product has a label, expect to fix the lettering afterwards rather than getting it right in the first pass. Generate a clean version with a believable label shape and blank or simple text, then composite the real typography in an editor. Trying to force a long product name through a prompt usually wastes a dozen generations.
A workflow that actually holds up
Write your prompt in three blocks and it becomes much easier to debug. First the scene and light. Second the product, described by material and finish rather than by brand. Third the camera language, meaning lens, depth of field, and angle. When a result looks wrong, you can usually trace it to one block and fix just that one instead of regenerating blindly.
Generate in batches of four and compare at thumbnail size. Product images live small on category pages and in ads, and problems that are invisible at 100% zoom are obvious when the image is the size of a postage stamp. If it doesn't read at thumbnail size, it doesn't work.
When to stop generating and start editing
No generation will be perfect. At some point the fastest path is to take a good-enough render into an editor and fix the two or three things that bother you: straighten the horizon, clone out a stray object, drop in real text, correct a white balance that drifted warm. Knowing when to switch saves hours. A render at 85 percent plus ten minutes of editing beats fifty more attempts at 100 percent.
Start with the light and the depth of field. Those two changes alone move most product renders from obviously artificial to plausibly real, and you can make them in the same tool you're already using. Head back to the generator and try one product with window light, a contact shadow, and an 85mm lens note. The difference shows up on the first batch.