Palette: teaching a small model to build decks in your design language
huggingface ·
IBM Research introduces Palette, a system that fine-tunes a 20B open-weight model to generate slide decks adhering to specific corporate design languages. To address the limitations of small models in handling complex layout geometry and editorial decisions, the approach splits the task into two roles: a designer who creates a structured JSON brief and a coder who generates the presentation code. The training data is synthesized using a large teacher model, incorporating counterfactual reasoning to teach the model why certain design choices were made over others. Users can steer the output with plain English directives wrapped in brackets, allowing for precise control over elements like charts or timelines. A critical component is a check-and-repair harness that validates every generated slide for overlaps or overflow, automatically fixing errors without degrading quality. In evaluations on PresentBench, Palette ranked first in Economics and Education domains and second overall among three measured categories. This method allows organizations to maintain ownership of confidential content while achieving visual consistency comparable to frontier AI systems.
huggingface