🔍 Read the full analysis: A Practical Guide To Teaching AI Your Design Language For Decks 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 describes Palette, a 20-billion-parameter open-weight model fine-tuned to generate slide 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 omits scores, benchmark details and information about a public release.
Hugging Face says it fine-tuned Palette, a 20-billion-parameter open-weight model, to create slide decks in IBM Design Language, reporting in the original analysis that it ranked first in two of three PresentBench domains it measured. The account describes a system that separates deck planning from slide coding, but supplies no scores, domain names or comparison details with which to assess the result independently.
Palette uses one LoRA adapter for two roles. Its “designer” role converts an input into a structured JSON brief covering the deck’s content, palette, typography, slide layouts and design rationale. A “coder” role then turns each slide specification into pptxgenjs JavaScript. Hugging Face says the split allows slide coding to happen in parallel, making that stage about six to seven times faster than running it sequentially. That figure is the company’s reported workflow comparison; the account does not provide test conditions or an independent measurement.
Because, according to Hugging Face, there is no public corpus pairing finished decks with their plans, design decisions and code, the team synthesized training examples using a larger open model. It says the dataset included about 1,000 decks and 10,000 slides, spanning different genres, lengths, layouts and input styles. Examples also included explanations of why a design choice was selected over an alternative. The post says coder examples were about five times longer than designer examples, so the team oversampled designer examples fivefold.
The team says it rendered generated decks to check for overlaps and text overflow, repaired broken slides and added revised examples to training. It also describes a factuality problem: when a specification left gaps, the model could invent material. According to the post, examples from reference decks helped guide the teacher model more reliably than adding further prompt rules. The account does not quantify how often these errors occurred or how frequently the checking process caught them.
A Structured Route to Branded Slides
Palette addresses two practical concerns for organizations considering AI-generated presentations: whether decks can follow an established visual identity and where sensitive input is processed. Hugging Face presents an open-weight model as a way for teams to gain more control over both. If a model can run on an organization’s own infrastructure, that may reduce reliance on an external service, but the project account does not report deployment results or establish that Palette protects confidential material in practice.
The design-and-code split also reflects a broader challenge in automating presentations. A deck requires editorial choices about what belongs on each slide as well as precise layout decisions about text, shapes and spacing. Hugging Face argues that separating those jobs gives the system a structured plan before it writes slide code. That approach could be useful to teams seeking editable, branded decks, but the reported benchmark result alone does not show how well it works across real-world presentation tasks.
The most defensible reading of the performance claim is narrow: Hugging Face reports strong results on part of its evaluation. Without scores, named domains, competing systems or test conditions, readers cannot tell how large the advantage was or whether it extends to other models and tasks.
Products Worth Considering
EcoFlow Delta 3 Portable Power Station – 1024 Wh, 1800 W
DJI Osmo Mobile 8 Smartphone Gimbal with Mic Mini Transmitter, Infinity Black
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
As an affiliate, we earn on qualifying purchases.
How Palette Plans and Builds Slides
The project account frames Palette as a response to shortcomings Hugging Face sees in general-purpose presentation generation. Such systems may fail to match an organization’s design system, while use of an external service may be unsuitable for confidential material. Those concerns are the motivation for the project, not reported evidence that Palette has already solved either problem.
Hugging Face says smaller models can make layout mistakes when they write slide code without evaluating the rendered result. The team’s workflow first creates a deck-wide brief and then generates code slide by slide. That brief can, according to the account, be edited by a person or another agent. The training process used LoRA fine-tuning, which the post says reduced the expense of experiments while adding the presentation format and skill to the base model.
The account focuses on IBM Design Language and says the method could be trained for a user’s design language. It does not provide results for other design systems. The distinction matters: a process described as adaptable is not the same as demonstrated performance across multiple brand systems.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s workflow
Benchmark and Release Details Missing
The supplied account does not name the PresentBench domains in which Palette ranked first, publish scores or identify the competing systems and evaluation conditions. It also does not identify the model used as Palette’s base, provide a model license, or include independent testing. Those omissions prevent a quantitative comparison with other systems.
It remains unclear how Palette performs on presentations involving complex data, strict factual requirements, long decks or design systems beyond IBM’s. The account does not show how reliably the checker detects layout failures, how frequently unsupported content appears, or what safeguards apply when the model is run on private infrastructure. A locally deployable model may give an organization more control over processing, but privacy and security outcomes are not established by the project description.
The post also does not say whether Palette’s weights or code are publicly available, when a release might occur, or whether further evaluation results are planned. Until those details are supplied, its benchmark rank and speed comparison remain claims reported by the project’s authors.
Awaiting Weights and Test Results
The next useful evidence would be a public release of model weights, code and evaluation details, allowing developers to reproduce the workflow and test it on their own decks. Hugging Face’s account gives no release date or confirmation that these materials will be published.
Further testing could clarify how the approach performs across different design languages, longer presentations and tasks where factual accuracy is particularly important. Results showing the rate of layout errors, unsupported additions and successful error detection would help organizations judge whether the workflow is suitable for production. For now, the documented training method and the company’s reported benchmark findings are available, but independent validation and deployment evidence remain absent from the supplied material.
Key Questions
What is Palette?
Palette is a 20-billion-parameter open-weight model that Hugging Face says it fine-tuned to generate slide decks in IBM Design Language.
How does Palette create a deck?
It first produces a structured design brief for the presentation, then writes pptxgenjs JavaScript to build slide layouts from that plan.
How did Palette perform on PresentBench?
Hugging Face reports that Palette ranked first in two of three measured domains. The supplied account does not provide scores, domain names, competitors or evaluation conditions.
Is Palette publicly available?
The supplied project account does not specify whether the model weights or code have been released, what license applies, or when a release may happen.
Does Palette keep company data private?
The project is motivated in part by the possibility of running an open-weight model on an organization’s own infrastructure. The account provides no deployment results or security assessment establishing how private data is handled in practice.
Primary source: Hugging Face · via ThorstenMeyerAI.com
