How to prevent your generated product images from looking identical
We spent weeks making lifestyle marketing generated product images with the default settings. Then we saw that every batch had the lighting angle and color palette. Our landing page suddenly looked like a stock photo collection.
The problem was that we used a default seed and fixed description words. To solve this we created a script that adds random camera focal lengths, different weather conditions and chosen film stocks into the prompt loop.
That small change made our generated product images ten times more varied. Gave our marketing a real editorial feel.
Generative models naturally go to the middle of their training data. You must push for diversity, in your parameter inputs if you want visual outputs.
Which parameters do you adjust most to break the generative look?
Replies
I went the other way from random parameters: every image gets its own brief written by a text model first. My app makes a cover for each generated song, and a cheap text call reads the lyrics and a caption of the finished track and writes a short brief for the image model: the subject, mood, palette, lighting and medium, photo, illustration or 3D. The image model never sees a template. Variety then comes from the input itself, a breakup ballad and a birthday song for a dog simply ask for different pictures. Random focal lengths fix the sameness, but they can also give a sad song a sunny beach. Diversity that is caused by the content stays on brand, diversity from dice does not always.