We're Training a Figma-Fluent Editing Model
Summary
Figma describes its approach to building a Figma-fluent editing model for on-canvas design edits, detailing supervised fine-tuning, synthetic prompts, and reinforcement learning. The post covers evaluation criteria, model comparisons, and how the Figma agent combines its own and third-party models to speed up edits while preserving quality and privacy. It highlights measurable gains in speed and user satisfaction and outlines future directions for training design models.