TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
IBM Research describes Palette, a system built around a fine-tuned 20-billion-parameter open-weight model that creates presentation decks using the IBM Design Language. The team says it ranked first in two of three PresentBench domains it measured; the post does not provide enough detail here to assess the size or significance of those results independently.
IBM Research has described Palette, a system that fine-tunes a 20-billion-parameter open-weight model to plan and generate presentation decks in the IBM Design Language. The team says the model ranked first in two of three PresentBench domains it measured, proposing a way for organizations to produce branded slides on infrastructure they control rather than relying on a hosted frontier model.
Palette divides deck creation into two roles handled by a single LoRA adapter. A designer role turns a user’s material into a structured JSON brief, specifying the visual system, slide content and layout, with a short rationale for each choice. A coder role takes each slide’s specification and writes JavaScript for the pptxgenjs library to render it. The separation is intended to keep editorial decisions—what a slide should say and how to present it—apart from geometric decisions such as where text and shapes fit.
The structured brief can be read and edited before slides are drawn, according to the report. Once the plan is set, the system can generate slide code in parallel; IBM Research says this makes the coding stage about six to seven times faster than running those calls sequentially. That figure applies to the coding stage described by the team, not necessarily to end-to-end deck production.
For training, the researchers generated about 1,000 decks covering roughly 10,000 slides, using a large open model as a teacher because they lacked a public dataset pairing design decisions with slide code. The examples include rationales that compare a chosen design with an alternative. IBM says every synthesized deck was rendered and checked for overlaps and text overflow; slides that failed were repaired and added back as training examples. The report also says the team oversampled designer examples fivefold because coder examples were about five times longer.
Branded Decks on Private Infrastructure
The project addresses two limits of using general-purpose models for corporate presentations: visual consistency and control over where confidential content is processed. The Hugging Face report argues that frontier models often return decks with a recurring house style, even when users want a particular template. A model that can run on an organization’s own infrastructure could offer more control over both the design language and sensitive source material.
Palette’s approach also reflects a practical trade-off. A smaller open-weight model may be more suitable for self-hosting, but the report describes common failures in slide generation, including overlapping elements, hidden titles and overflowing text. Separating planning from slide construction and checking rendered output aims to make those failures easier to catch. The claimed benchmark result is promising, but the supplied report does not establish that Palette performs better across all tasks, design systems or real-world workflows.
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
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Why Slide Generation Needs Two Roles
Creating a deck requires more than writing concise text. A system must decide what belongs on each slide, choose a visual form such as a chart or table, and place every element at specific coordinates and sizes. In the method described by IBM Research, the model writes code rather than directly seeing and adjusting a finished slide as a human designer might. Errors in estimating text length or placement may only become visible after rendering.
The team’s central design choice is to make the work resemble a staged human workflow: settle the story and slide specifications first, then build each slide. This also responds to a data gap identified in the post. Finished presentations are available, but they generally do not record the reasoning behind design choices or the code that produced the slides. Palette’s synthetic examples attempt to supply those missing intermediate steps and teach not just a layout, but why one option was selected over another.
“We fine-tuned a 20B model to build decks the way people do: a designer plans the deck and a coder draws each slide.”
— IBM Research, in the Palette report
Benchmark Scope and Deployment Limits
The supplied report does not include the PresentBench scores, comparison models, evaluation protocol or names of the three domains, so readers cannot determine how large the reported advantage was or how it compares with other systems. It also does not establish whether Palette’s results transfer beyond the IBM Design Language, or how well the model handles other organizations’ templates and brand rules.
Other practical details remain unspecified in the material provided, including the amount of editing decks need after generation, performance on confidential or highly variable source documents, and the cost and hardware requirements of self-hosting the 20-billion-parameter model. The report describes synthetic training and a rendering checker, but does not show that these measures eliminate invented facts or layout errors. IBM’s account of the reference-deck method likewise does not quantify how often it prevents factual additions.
Further Testing Across Design Systems
The immediate next step for readers evaluating Palette is to examine the full benchmark methodology and test the system on their own material, templates and infrastructure. The report outlines a workflow and its reported results, but the supplied source does not announce a release date, public model checkpoint or product rollout. Availability and broader independent evaluations are not established in the material.
For the approach to prove useful beyond the reported experiment, future evidence would need to show how reliably it preserves source facts, how much human correction it requires, and whether its checking process catches errors across different deck styles. Those results would help organizations judge whether a self-hosted model can meet their requirements for design control and data handling.
Key Questions
What is Palette?
Palette is a deck-generation system described by IBM Research. It uses a fine-tuned open-weight model to plan presentation content and create slides in the IBM Design Language.
How does Palette build a presentation?
One role writes a structured deck brief with slide content, layout and design rationale. A second role generates JavaScript to draw each slide, and a checker examines rendered decks for issues such as overlap and overflow.
What benchmark result did IBM report?
IBM Research says Palette ranked first on two of the three PresentBench domains it measured. The supplied material does not state scores, comparison systems or the evaluation details needed to interpret the result more fully.
Can organizations use Palette with their own design language?
The report presents training a model on an organization’s design language as the broader goal, but the system described was trained on the IBM Design Language. The supplied material does not confirm a public process or release for adapting Palette to other brands.
Is the model publicly available?
The supplied report describes the model and training method but does not confirm that a model checkpoint, product or deployment package is available to the public.
Source: huggingface.co
