Google unveils PaLM Canvas: a multiturn, multimodal model aimed at rapid prototyping

AI · 6 min read

PaLM Canvas supports multiturn interactions where users can progressively refine a screen, ask the model to alter layout constraints, or request accessibility adjustments. It keeps a persistent contextual state, which means designers can say "make the menu more compact" and the model will apply the change across all relevant frames in the session.

Google highlighted Canvas's multimodal fidelity: it can output layout JSON, annotated SVGs, and copy variations while maintaining internal consistency across steps. The model integrates with Google Workspace and Figma plugins to let teams capture Canvas sessions as artifacts and export final assets directly to project files.

Privacy and control are central to the product's positioning: Google offers Workspace admins options to limit cloud retention of Canvas sessions and supports enterprise-only models via the company's cloud containment solutions. The company is also publishing a design-oriented benchmark to audit Canvas's layout consistency and accessibility improvements.