Figma brings back its AI design generator as First Draft

Figma is giving its AI generator a second chance.

Back in June, the company introduced Make Designs, a tool meant to help designers jumpstart their workflows. But early users quickly noticed a glaring issue: asking it to design a weather app produced an interface that looked almost identical to Apple's native iOS weather app. Figma insisted it never trained the underlying models on customer data, but pulled the tool to investigate.

Now it is back under a new moniker, First Draft, rolling out in a limited beta after what the company calls extensive analysis and testing.

Inside the new engine

To clear up the confusion, Figma detailed exactly how the feature operates under the hood.

First Draft does not train on customer designs. Instead, it relies on off-the-shelf foundation models, including OpenAI's GPT-4 and Amazon Titan, and couples them with Figma's own context.

That context is basically a proprietary design system packed with mobile and desktop UI components and layout examples. When you type in a prompt, the AI does not paint a picture from scratch. It acts like an assembler, picking out pre-built components and arranging them into an editable layout.

Figma is also introducing four distinct libraries for users to choose from. That includes a lo-fi wireframing library for sketching out basic app logic without heavy visual styling, alongside higher-fidelity options for visual exploration.

Why it matters

Figma's pivot here is remarkably pragmatic.

For product designers, pure generative image models have always been a dead end because you cannot easily edit layers, resize autolayouts, or swap out components. By constraining the AI to select and arrange actual design system components, Figma gives teams an editable starting point rather than a rigid graphic. The new low-fidelity wireframe option also feels like a direct fix for the original problem: it stops the tool from generating overly styled mockups that accidentally mimic famous apps.

Across the industry, this sets a practical playbook for generative UI. Instead of trying to train models to reinvent app design, the winning move looks like using existing LLMs as orchestrators of structured design tokens. It keeps copyright risks low and utility high.