# IBM Research introduces Palette: a small model for custom design language slide generation

> **Key takeaway:** Palette uses a dual-role LoRA adapter to teach small 20B models how to generate structured slide decks using specific design languages and JavaScript code.

**Published:** 2026-11-03T16:35:36+00:00 
**Source:** Hugging Face Blog & Announcements 
**Category:** ai-open-models 
**Canonical URL:** https://fosswire.org/news/ibm-research-introduces-palette-a-small-model-for-custom-des.html 

## Summary
IBM Research has unveiled Palette, a specialized approach for training small-scale open-weight models to build professional slide decks that adhere to specific design languages. Unlike frontier models that often default to a generic 'AI-generated' aesthetic, Palette is designed to be self-hosted on private infrastructure to maintain data confidentiality and brand consistency.

The system utilizes a 20B parameter base model enhanced with a LoRA adapter. It solves the common failure modes of small models—such as overlapping text or broken layouts—by splitting the deck-building process into two distinct roles: a 'designer' that creates a structured JSON brief and a 'coder' that generates the necessary JavaScript (pptxgenjs) to render the slides.

## The Designer-Coder architecture and synthetic training
- **Dual-Role LoRA Adapter:** The model splits tasks into an editorial 'designer' role for content planning and a geometric 'coder' role for precise element placement.
- **Synthetic Data Pipeline:** Training data was generated using a large open teacher model to create approximately 1,000 decks and 10,000 slides, including counterfactual reasoning to explain design choices.
- **Automated Quality Harness:** A rendering checker identifies layout errors like text overflow or shape overlaps, feeding these mistakes back into the training loop as repaired examples.
- **Parallelized Execution:** By separating the JSON design brief from the slide-specific code, the coding stage can be executed in parallel for each slide, increasing generation speed by 6–7×.
- **Directives for Control:** Users can steer specific design elements using plain English instructions wrapped in double brackets within the rollout plan.

## Why it matters
This development provides a blueprint for enterprises to deploy highly specialized, small-scale AI models that respect strict design standards and data privacy. By moving away from massive frontier models toward fine-tuned 20B models, organizations can achieve professional-grade automation on their own hardware without vendor lock-in.