IBM's 20B model uses synthetic data to rank first for consistent branded slide decks on PresentBench

IBM Research published a blog post detailing how a 20B parameter model can build slide decks that adhere to specific design languages.confirmed The model ranks first on two of the three domains measured in the PresentBench evaluation, according to the post.reportedIBM Research shows a small model can create branded slide decks that look consistent with company templates.A 20B LoRA-fine-tuned model achieves top PresentBench scores by separating editorial planning from geometric coding.

The evidence

The team fine-tuned a base model using a Low-Rank Adaptation (LoRA) adapter rather than full fine-tuning. This approach allowed them to teach a specific format and skill without disturbing the base model's existing knowledge. The training data consisted of approximately 1,000 decks covering roughly 10,000 slides, generated by a large open teacher model.reported Recipes set the genre, length and layouts for each deck, ensuring controlled coverage of the data. Plans ranged from tidy outlines to raw content dumps, forcing the model to learn condensation and cutting.

The decision

Palette splits deck creation into two roles served by a single LoRA adapter: a designer and a coder. The designer reads a plan and writes a structured JSON brief containing palette, type system and layout details for each slide. The coder then generates code for each slide in parallel, making the coding stage about six to seven times faster. This separation allows the designer to handle editorial judgment while the coder handles geometry, reducing errors in placement and sizing.

The takeaway

Small open models often fail at deck creation because they cannot see the slide while laying it out. They write code placing elements at specific coordinates but only discover errors after rendering, leading to overlapping shapes or text running off slides. Palette addresses this by separating editorial decisions from geometric placement. The JSON brief between roles is readable and editable, allowing humans or agents to modify a single slide's layout without regenerating the entire deck.

Building on this

Developers can build self-hosted slide generation tools that maintain brand consistency using the documented two-stage architecture. The separation of designer and coder roles enables parallel processing, reducing generation time for multi-slide decks. Because the intermediate JSON brief is editable, developers can insert human review steps or agent-based corrections before final rendering. The model supports open weights, allowing enterprises to keep confidential content within their own infrastructure. Training on synthetic data demonstrates that small models can learn complex design languages without public corpora of design rationale.