Palette: teaching a small model to build decks in your design language
Blog post from Hugging Face
Palette is an IBM Research approach for generating PowerPoint decks with a self-hosted 20B open-weight model trained to follow an organization’s design language, addressing the tendency of frontier models to impose a generic visual style and the need to keep confidential content on private infrastructure. It divides deck creation into a designer role that produces an editable JSON brief covering narrative, layouts, typography, and rationale, and a coder role that generates PPTXGenJS for individual slides in parallel, improving generation speed and separating editorial decisions from geometric placement. Because no suitable public training corpus existed, the team synthesized about 1,000 decks and 10,000 slides using a larger teacher model, including counterfactual reasoning, reference examples, user directives, and iterative repairs for rendering errors. A check-and-repair harness renders slides, detects clipping, overflow, overlap, and other defects, and accepts editor-model revisions only when measurable checks confirm improvement, increasing PresentBench scores by an average of 3.5 points across 131 decks. Palette ranked first on the Economics and Education PresentBench domains, third on Talk, and second overall behind NotebookLM, while future work includes using automated checks as reinforcement-learning rewards, improving figure-heavy decks, and adapting output to existing PowerPoint templates without retraining.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| AI Model Fine-tuning | 3 | 613 | 130 | 56 | -29% |
| Reinforcement learning | 1 | 99 | 55 | 28 | -39% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.