FaV-A generates positive-negative-space artwork through three stages: creating a base object, analyzing its contours and spatial structure for secondary meanings, and turning those findings into instructions for final image synthesis.
Qualitative comparisons, user evaluation, and ablations indicate that this staged process yields more structurally coherent and semantically aligned compositions than direct zero-shot multimodal prompting. The approach is interesting because it separates structural discovery from final rendering instead of asking one generation pass to solve shared-boundary geometry and dual semantics simultaneously.