Devlog · Art & Concept
Land and air support vehicles needed concept art before anything else about them was locked. Generated independently, all six looked like they belonged to six different games. Here's what fixed it.
Before a vehicle has a stat block, it has a silhouette. Six support vehicles — three land, three air, including the Iron-Hail bomber pictured above — needed concept art before their combat roles were even fully locked, and the first pass exposed a problem that has nothing to do with balance.
Generated independently, one prompt per vehicle, the first pass came back looking like six different concept artists had never spoken to each other. Different material language, different grime levels, different sense of scale. Individually, several of the renders were strong. As a set, they didn't read as one army.
The actual issue: a shared style description in a text prompt isn't enough to guarantee a shared visual identity. Two prompts can both say "steampunk, weathered iron, brass fittings" and still produce vehicles that look like they're from unrelated settings.
The fix was to stop treating style as something described and start treating it as something inherited. Generate the strongest hero render first, then carry a direct visual reference from that image into every other prompt in the set — not a written description of it, the image itself as a style anchor.
Everything after that first hero shot locked to the same reference: the same weathering level, the same palette, the same sense of how brass catches light. Six separate generations, one shared DNA.
Production note: this is now the standing rule for any multi-unit concept art set — generate the anchor first, pick the strongest result, then seed every subsequent prompt from that specific image rather than re-describing the style from scratch each time.
With individual units consistent, the second problem was making land and air read as distinct at a glance without breaking the shared style. Two additional composition pieces staged each three-unit set together in a single frame — one land trio, one air trio — specifically to check whether the sets were readable as a group, not just handsome one at a time.
Consistency isn't a filter you apply after the fact. It's a decision you make before the first image generates.
← All Devlog Posts