From the source
IBM's research team trained Palette, a 20B open-weight model that builds slide decks in an organisation's own design language, splitting the work into a designer that writes a structured brief and a coder that draws each slide.
- Small self-hosted models usually produce broken slides: titles that vanish into the background, text running off the slide, shapes stacked on top of each other.
- Palette was trained with a LoRA adapter rather than full fine-tuning, and its training data follows the IBM Design Language.
- On PresentBench it ranks first on two of the three domains the team measured.
- The training data is synthetic: roughly 1000 decks and about 10,000 slides, generated with a large open teacher from recipes that set genre, length and layouts, and each example carries counterfactual reasoning for why an alternative was rejected.
- Every deck is rendered and checked for overlaps and overflow by the same checker used at inference, and broken slides are repaired with their reasoning rewritten, so each caught mistake becomes a training example.
- Because the designer fixes every slide's context before any code is written, the coder calls run in parallel, making the coding stage about 6–7× faster.
