Fathom Labs unveils CanvasGen: multimodal model that generates editable Figma layers

AI · 6 min read

CanvasGen accepts prompts combining text, sketches, and reference frames and returns a serialized Figma layer tree with named components, constraints, and tokenized styles. The model is explicitly trained to respect design system tokens and include metadata for accessibility and localization.

Fathom Labs provides a direct Figma plugin and an API that returns delta updates suitable for patch-based collaboration. The company says the model reduces manual cleanup after generation because it prioritizes component reusability and semantic naming conventions.

Privacy-conscious teams can run CanvasGen on self-hosted infrastructure; Fathom Labs offers an enterprise deployment with model fine-tuning on private design systems. Early adopter case studies show reductions in mockup-to-prototype time, particularly for complex responsive layouts.