Description
TK Events is building **Virtual Creator Studio (VCS)**, a browser-based platform where users design their own virtual event venue by describing it in plain language. Our AI editor generates a photorealistic image of each room (lobby, main stage, expo hall, resource room), and our app overlays interactive UI on top: room-navigation cards, sponsor banners, livestream displays, exhibitor booths. We currently generate images through Google's nano banana (Gemini 3.1 Flash Image) using a layered prompt system: Global rules + Room-specific rules + a separate Mapping prompt that returns JSON coordinates for our overlay layer. ## The problem The AI-generated image and our overlay coordinates don't line up reliably. Specifically: - The lobby's central screen shifts by 10–20% in position and size across generations, even with locked coordinates in the prompt. - Room-navigation cards duplicate, drift in count, or render generic labels ("Room 1", "Section A") instead of the actual room names passed in via variables. - Sponsor banner zones aren't consistent between generations, so imagery overlays don't sit cleanly. - Canvas aspect ratio (16:7.9) isn't honoured — the model outputs at its own preferred ratio, causing visible distortion when displayed in our 16:7.9 container. Text-only prompting has hit its ceiling. We need a specialist to redesign the pipeline. ## What we need you to deliver A working generation pipeline that meets the following bar: 1. Screen and banner zones land within ±