Meta Publishes Llama-4 CodeArt: Multimodal Model for UI Code Generation

Tech · 6 min read

Meta's Llama-4 CodeArt introduces a multimodal architecture trained specifically to translate visual designs into idiomatic UI code. It accepts screenshots, layered design files, or a combination of both and outputs componentized code with CSS-in-JS or platform-equivalent styling.

To improve fidelity, CodeArt includes a verification pass that simulates rendering and compares the visual output to the source image; mismatches are flagged and an annotated diff is returned to the developer or designer. The model also generates accessibility metadata and test IDs where applicable.

Meta stressed responsible deployment and released a developer policy for code generation, including recommended human-in-the-loop checks and open-source dataset provenance notes. Early adopter companies reported significant time savings converting high-fidelity mocks into scaffolded apps.