A library of everyday Belgian scenes for a pay and benefits platform: people at work, on the move, on holiday, at the checkout. Shot like flash-lit documentary photography, with faces and places that read as real rather than generated.
Why one model wasn't enough
No single model made these images.
I used GPT Image 2 to build factually grounded base images, Soul 2.0 to translate them into a more editorial visual language, and Seedream 5 / Nano Banana 2 to repair spatial logic, object interactions, and difficult compositions. When scenes required coherent text or graphic elements, I returned to GPT Image 2 for additional edits. The result was a flexible workflow where images moved between models according to their strengths rather than being forced through a single pipeline.
Real-world logic, place knowledge, composition, text/graphics, scene setup
Editorial tone, more interesting image character, less generic AI faces
Spatial fixes, object relationships, hard compositions, gesture cleanup
Branding, text, interface elements, factual or graphic corrections
Not linear — images could skip steps, loop, or bounce backward depending on the shot
