LumaAI unveils Palette-2: a 12B visual design foundation model for mockups

AI · 5 min read

Palette-2 is optimized for visual design workflows, combining layout understanding, typography inference, and asset generation in a single model. LumaAI says it was trained on a mix of licensed design files, synthetic UIs, and public web screenshots to balance fidelity with generalization. The model can accept a hand-sketched wireframe and produce layered, export-ready mockups in Figma and SVG formats.

A notable capability is style anchoring: designers can lock a color palette, type scale, or brand tokens and Palette-2 will re-skin generated screens to conform. The model also returns structured layout metadata so teams can hook outputs into design systems and automated QA pipelines. LumaAI ships a web demo and an API with rate limits tuned for collaborative design sessions.

Privacy and IP remain a focus. LumaAI offers an enterprise on-prem option and a redact mode that strips examples from model telemetry. Early reviewers praise the speed and editability of outputs but note occasional hallucinated text and accessibility oversights, which the company says will improve with fine-tuning and community feedback.