🔍 Read the full analysis: Build Decks In Your Design Language With Palette’s Small Model on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Hugging Face says it fine-tuned a 20-billion-parameter open-weight model, Palette, to generate presentation decks in IBM Design Language. The company reports top rankings in two of three PresentBench domains and a faster slide-coding workflow, but the supplied account includes no scores, domain names, or independent evaluation.
Hugging Face says it fine-tuned Palette, a 20-billion-parameter open-weight model, to create slide decks in IBM Design Language. The project combines a planning stage with code generation for individual slides; the company reports Palette ranked first in two of three PresentBench domains it measured, but has not provided the scores or evaluation details in the source material.
Palette uses a single LoRA adapter for two roles. Its “designer” role turns a request into a structured JSON brief covering the deck’s content, palette, typography, layouts and design rationale. A “coder” role then uses specifications for each slide to write JavaScript with pptxgenjs, the library used to generate the presentation file.
Hugging Face says this division lets the slide-coding stage run in parallel, making it about six to seven times faster than sequential coding. That is a project-reported speed comparison; the supplied account does not give test conditions or clarify whether the figure applies to total deck-generation time or only the coding stage.
The team says it synthesized about 1,000 decks, or roughly 10,000 slides, for training because there is no public corpus containing finished decks alongside their plans, design choices and code. The examples covered different genres, lengths, layouts and input styles. Hugging Face also says it rendered generated decks to check for overlapping elements and overflowing text, repaired broken slides and added examples incorporating revised design reasoning.
A Workflow for Branded Decks
Presentation generation involves both deciding what a deck should say and fitting that material into slides. A model that produces plausible text can still make slides difficult to read if text overflows, elements overlap or the layout does not follow an organization’s visual system. Palette’s planning-before-coding workflow addresses that problem by establishing a deck-wide brief before producing slide code.
The project also speaks to organizations that want presentations to follow their own design rules or have concerns about sending confidential material to an external service. Hugging Face presents open weights and the possibility of running a model on an organization’s infrastructure as potential benefits. Those are motivations for the work, however, not demonstrated deployment outcomes: the supplied account does not establish that Palette has been deployed privately or show how it handles sensitive information.
The reported benchmark result is a limited signal that a smaller model may perform well on some presentation tasks. Without scores, named domains, competing systems and test conditions, readers cannot tell how large the advantage was or compare Palette’s performance across the broader field.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
Samsung 34-Inch Odyssey G5 Ultra-Wide Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Palette Was Trained
Hugging Face says the training data was created with help from a larger open model because examples linking presentation plans, design decisions and slide code were not publicly available. The synthetic examples included counterfactual reasoning—explanations of why one design choice was selected instead of an alternative. The team says reference-deck examples helped guide the teacher model when it produced unsupported content in response to underspecified briefs.
The source also describes a data-balance issue: coder examples were about five times longer than designer examples. Hugging Face says it oversampled designer examples fivefold to give that role more representation during training. LoRA was used to adapt the base model, which the account says reduced the cost of experiments while adding presentation-related skills and formats.
Hugging Face frames the two-role setup as a way to separate editorial decisions from slide geometry. The JSON brief can be edited by a person or another agent before code is generated, offering a potential review point between planning and production. The source does not report how often users make such edits or how much they improve the final decks.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s approach
Benchmark and Release Details
The supplied account does not identify the PresentBench domains in which Palette ranked first, provide scores, name the competing systems or explain the evaluation conditions. The ranking is therefore a claim from the project’s authors that cannot be independently assessed from the available details. It also does not show independent testing of the resulting decks.
The source does not name the model’s 20B base model or state whether Palette’s weights, code or evaluation materials are publicly available, under what license, or when they might be released. Performance on complex data, strict factual tasks, longer decks and design languages beyond IBM’s is also unreported. Hugging Face describes a method that could be trained for a user’s design language, but the supplied material gives results only for IBM Design Language.
Another open issue is factual reliability. Hugging Face acknowledges that the model could invent material when a brief lacked information and says reference-deck examples helped address that behavior. The source does not quantify how often unsupported content appeared after this change or how reliably the checking process catches layout problems.
Evidence Needed for Wider Use
The next useful details would be publication of Palette’s weights, license, benchmark scores and testing methodology. Those would let developers check the reported rankings and assess the approach on their own material. No release date or schedule for these materials is given in the supplied account, so their availability remains unknown.
Further testing would need to show how the workflow performs with other design systems, longer or more demanding decks, complex data and strict factual requirements. Measures of text overflow, slide overlap and unsupported claims would help establish how well its automated checks work in practice. Until such evidence is available, the speed figure and PresentBench rankings should be read as claims reported by the project team, not as independently verified results.
Key Questions
What is Palette?
Palette is a 20B open-weight model that Hugging Face says it fine-tuned to create slide decks in IBM Design Language. Its workflow produces a structured deck brief before generating code for individual slides.
What did Hugging Face report about its benchmark results?
The company says Palette ranked first in two of three PresentBench domains it measured. The supplied material does not name those domains or provide scores, competitors or test conditions, so the result cannot be assessed independently from those details.
How does Palette generate a deck?
A designer role creates a JSON plan covering content and visual choices. A coder role then uses slide specifications to generate pptxgenjs JavaScript for the presentation.
Is Palette available to download?
The supplied account does not say whether the model weights or code have been released, identify a license or give an availability date. Release status is unclear.
Does Palette support design systems beyond IBM’s?
Hugging Face presents the workflow as adaptable to a user’s design language, but the supplied material reports on IBM Design Language and gives no results for other systems.
Primary source: Hugging Face · via ThorstenMeyerAI.com
