A generative design pipeline for bridal wear
Three demos showing how a single multimodal model (Gemini 3) can act as the engine for an end-to-end design tool: editing traditional motif stencils, rendering them onto silk garments, and dressing the finished garment on a model — with the original design preserved at every step.
Demo 1
Pick a traditional khaka stencil. The model describes it, swaps a crucial design element on instruction, and regenerates the motif — heraldic fleur-de-lis becomes a lotus, scrollwork preserved.
Demo 2
A flat designer illustration of a lehenga with placed motifs is rendered first as a photorealistic silk garment, then as a lookbook shot of a model wearing it. Every placed motif is preserved pixel-faithful; only drape and lighting change.
Demo 3
Two khaka silhouettes × two embroidery techniques (Phulkari and Zardozi). Same shape, four surface treatments. Demonstrates precise control over what changes (color, texture) and what doesn't (outline, proportion).
What this demonstrates
The pipeline is a chain of constrained image-to-image transformations through one multimodal model. Each stage's output feeds the next, so the final result still carries the original design language — peacock plumage, marigold flowers, side-panel butis — but now warped to the body and lit like real silk.
The technical core is prompt design: each Gemini call explicitly tags its inputs by role ("this is the protected design", "this is a paint-chip swatch — sample its colour only") and encodes preservation rules as anti-pattern failure clauses. Where prompting wasn't enough, a small preprocessing step (e.g. padding a swatch to a square canvas) closed the remaining gap.