Product photos generated with AI often look great once and then never again. The fix is not a better single prompt, it is a template you can reuse across every SKU so results stay consistent. Here is the structure I use.
1. Separate the fixed parts from the variable parts
Split your prompt into two layers. The fixed layer describes your brand look: background, lighting, camera angle, and mood. The variable layer describes only the product: its name, material, color, and size. When you change products, you only touch the variable layer, so every image in the catalog shares the same visual DNA.
A simple skeleton looks like this:
[PRODUCT]: matte ceramic mug, sage green, 350 ml
[SCENE]: seamless warm-gray backdrop, soft shadow under object
[LIGHT]: large softbox from upper left, gentle fill from right
[CAMERA]: 50mm, eye level, centered, shallow depth of field
[STYLE]: clean e-commerce photography, true-to-life color
2. Be specific about materials
Models guess wrong when you write "nice bottle." Write "frosted glass bottle with brushed aluminum cap." Material words control reflections, texture, and highlights more than any style keyword does.
3. Lock the composition
Decide once whether products sit centered or on a rule-of-thirds line, and how much empty space surrounds them. Write it into the template. Consistent negative space makes a product grid look professional even when the items are very different.
4. Keep a negative list
Maintain a short list of things you never want: extra props, text artifacts, warped logos, duplicated objects. Append it to every prompt rather than fixing problems after the fact.
5. Version your template
Store the template in a text file or a repo next to your assets. When you tweak the lighting line, bump a version number and note why. Later, when someone asks why the spring collection looks different, you will have an answer.
6. Test with three products before scaling
Run the template on a light item, a dark item, and a reflective item. If all three look right, the template is ready for the full catalog.
I have been testing this workflow with Nano Banana 2, and the layered structure made batch results noticeably more consistent. Whatever model you use, the template is what turns lucky shots into a repeatable process.
Top comments (0)