Palette: teaching a small model to build decks in your design language
Hugging Face introduces Palette, a 20B open-weight model fine-tuned to generate slide decks adhering to a user’s design language while remaining self-hostable.
Frontier models produce generic slide decks that lack custom design language, and open-weight models often fail to render slides correctly. Palette addresses this by fine-tuning a 20B model to generate decks in a specified design language, tested against the IBM Design Language on PresentBench where it ranked first in two of three domains.
Small open models struggle with slide layout, as they must calculate precise positioning and rendering without visual feedback. Palette splits the task into two roles—designer for editorial judgment and coder for geometry—using a single LoRA adapter, enabling parallel slide generation and reducing coding time by 6–7×.
Training data for Palette was synthesized using a large open teacher, generating around 1,000 decks with 10,000 slides. Each example includes reasoning in counterfactual form to teach design rules, and a harness renders and checks every slide for errors before output.
The model supports user directives embedded in the plan to enforce specific design choices, such as forcing a timeline instead of a table. A harness repairs errors like overlapping elements or incorrect colors, ensuring the final deck meets quality standards before delivery.