It is generative. I have used a combination of LLM, fined tuned text to image models and a model that tags the components from design systems. Post this I am writing code to convert this to figma components using figma API. Also using some heuristic algorithms to correct for layout artefacts.