Palette: teaching a small model to build decks in your design language
What it is
An IBM Research blog post describing Palette, a method that fine-tunes a 20B open-weight model (via a single LoRA adapter) to generate slide decks in a chosen design language. It splits the task into two roles: a 'designer' that writes a structured JSON brief (palette, type system, per-slide layout, content and rationale) and a 'coder' that emits pptxgenjs JavaScript to draw each slide, trained on synthetic data following the IBM Design Language and wrapped in a check-and-repair harness. The post reports that on PresentBench the model ranks first on two of the three domains measured.
The problem it solves
Frontier models produce decks with a recognizable generic 'AI-generated' house style that prompting cannot steer to a specific template, and small self-hostable open models produce broken layouts — titles lost in the background, text running off the slide, overlapping shapes.
What you could use it for
- Generating on-brand slide decks that follow a company's own template and design language
- Self-hosted deck generation for confidential content that cannot be sent to a frontier API
- Producing editable PowerPoint output as pptxgenjs code rather than flat images
- Research on synthetic data generation and check-and-repair harnesses for layout/geometry tasks
Why it matters
If the approach holds up, enterprises could own and self-host a small model that matches their design language instead of relying on frontier models that drift back to a generic house style, while keeping confidential content inside their own infrastructure.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.