Skip to content
Tools · Nov 3, 2026

IBM Research releases Palette, a fine-tuned 20B model for generating slide decks in a user’s design language

Palette uses a two-role LoRA adapter—designer and coder—to plan and render slides, and ranks first on two of three domains in PresentBench.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • IBM Research fine-tuned a 20B open-weight model to generate slide decks that adhere to a user’s design language and avoid the generic “AI look.”
  • The model splits work into a designer that writes a JSON brief and a coder that writes JavaScript to draw each slide, enabling parallel rendering and easier edits.
  • Training relied on roughly 1,000 synthetic decks (10,000 slides) with counterfactual reasoning traces; a check-and-repair harness fixed errors before they entered the training set.
  • On PresentBench, the fine-tuned model ranked first on two of three domains; the authors attribute gains to reference decks and a repair loop rather than prompt rules.

IBM Research describes Palette, a fine-tuned 20B model that generates slide decks adhering to a user’s design language and avoiding the generic “AI look” seen with frontier models. The authors argue that frontier systems drift toward their own house style and cannot be reliably steered to match a proprietary template, and that small open models often produce broken layouts when asked to render decks on their own.

Sources
  1. 01Hugging Face — Palette: teaching a small model to build decks in your design language
Also on Tools

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.