Gempix Blog
Notes from the team on prompts, models, design decisions, and the occasional opinion about where this is all going.
Most people spend five seconds on a prompt and five minutes choosing between styles. We looked at hundreds of generations and found the opposite approach produces better results — almost every time.
The majority of visual content is consumed on phones in portrait orientation. Yet most design tools default to landscape. Here is how Gempix thinks about format — and why 864 × 1152 is not an arbitrary choice.
Keywords get you images. Language gets you art. The difference is in what gets lost when a model processes your input — and what we are doing to preserve it.
We tested over 200 style descriptors before settling on the current library. Some were technically impressive but creatively limiting. Others were popular but redundant. Here is what the curation process looked like.
Watermarks on AI-generated images started as a transparency mechanism. They have mostly become a revenue gate. We thought hard about where we stand on this, and why.
A desktop hero, a story card, and a social square have almost nothing in common dimensionally. Making a single prompt work across all three required rethinking what "composition" means when the canvas keeps changing.
No frequency commitment. We write when we have something worth saying.