IBM Research has introduced Palette, a system for generating slide decks in a specific design language using a fine-tuned 20-billion-parameter open-weight model. Trained on synthetic examples based on IBM Design Language, Palette separates planning each deck from writing the code that renders its slides, then checks and repairs the output before returning it.
The project addresses two limitations of using general-purpose frontier models for presentations: their output can fall back on a familiar house style, and confidential content may need to remain on a user’s own infrastructure. Smaller open models can be self-hosted, but often produce basic layout errors, including text that runs off a slide or elements that overlap.
Palette divides deck creation between two roles served by one LoRA adapter. A “designer” turns a plan into a structured brief covering the visual system, slide layouts, content and design rationale. A “coder” then writes the JavaScript for each slide using pptxgenjs. Separating editorial decisions from geometry also allows the slide-generation calls to run in parallel; the team says this makes that stage about six to seven times faster.
Because there is no public collection pairing presentation plans with design decisions and slide code, the researchers generated around 1,000 decks covering roughly 10,000 slides with help from a large open model. The examples span different genres, lengths and layouts, and include plans ranging from polished outlines to raw content dumps. They also include reasoning that explains why a design choice was made and why an alternative was rejected.
The training pipeline renders each synthetic deck and checks for problems such as overlap and overflow. Broken slides are repaired, and the corrected examples are added to the training data with revised reasoning. The researchers found that this process improved the examples more reliably than adding prompt rules when the teacher model made mistakes. Reference decks also helped guide its output toward the intended design language.
Users can add short instructions, called directives, to control specific choices. Examples include asking for a timeline instead of a table, specifying the audience, changing the tone or emphasizing a key figure. The designer incorporates each directive into the brief, and the coder uses it when generating the relevant slide.
Palette does not return generated slides without checks. Its harness first applies simple fixes, then renders and measures each slide to detect overlapping, clipped or off-slide text. Slides that fail are sent to a separate editor model for rewriting. A rewrite is accepted only if the checker confirms that it improved the slide; otherwise, the original is kept. In tests on 131 PresentBench decks, the harness increased scores by an average of 3.5 points.
On PresentBench, which assesses fundamentals, visual design, completeness, correctness and fidelity, Palette ranked first in the Economics and Education domains, and third in Talk. Its average across those three domains was second, with a score of 59.7 compared with 60.8 for NotebookLM. The evaluation used a separate frontier model to turn each source document into a plan before Palette built the deck.
The researchers note that PresentBench evaluates deck quality but does not measure whether a presentation follows a particular design language. They say future work could use the harness’s checks as rewards for reinforcement learning, extract figures from source documents for figure-heavy decks, and explore adapting output to existing PowerPoint templates without retraining the model.
Sources: Hugging Face Blog.