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 project that fine-tunes a 20-billion-parameter open-weight model to create slide decks in the IBM Design Language. The team says it ranked first in two of three measured PresentBench domains; the report does not provide enough detail here to assess those results independently or establish how well the approach transfers to other design systems.
IBM Research has reported Palette, a system that fine-tunes a 20-billion-parameter open-weight model to plan and generate slide decks in the IBM Design Language. The team says the model ranked first in two of three PresentBench domains it measured, presenting the project as a way to produce branded decks on infrastructure an organization controls.
Palette divides deck creation into two stages, rather than asking a model to plan and draw every slide in one pass. A designer role creates a structured JSON brief with the deck’s palette, typography, slide content and layout choices. A coder role turns each slide specification into JavaScript using pptxgenjs. Both roles use a single LoRA adapter, a fine-tuning method that adds task-specific behavior to a base model.
The team says the split separates editorial decisions—what a slide should communicate and whether it needs a chart or table—from geometric work, such as sizing text and placing objects. Since the brief fixes the slide plans before code generation begins, slide-coding calls can run in parallel. The report estimates that this makes the coding stage about six to seven times faster; it does not provide a broader end-to-end timing comparison in the supplied material.
To create training data, the researchers used a large open model to generate about 1,000 decks covering roughly 10,000 slides. The examples included planning and design rationales as well as code. The team rendered the slides, checked for overlaps and text overflow, and repaired broken examples. It also says it oversampled designer examples fivefold because coder examples were roughly five times longer, a difference that otherwise left the designer role undertrained.
Branded Decks on Private Infrastructure
Palette addresses two practical limits of general-purpose deck generation raised by its authors: models can produce decks that resemble one another, and confidential material may not be suitable for external services. A self-hosted open-weight model could give organizations more control over where content is processed and how generated slides look. That matters to teams with established templates, internal data or review requirements.
The reported workflow also makes some design decisions easier to inspect and edit. Its intermediate JSON brief can be changed without regenerating the whole deck, according to the report. That offers a potential route for human oversight: a person can revise a slide’s content or layout before code is produced. The project does not, however, establish that every organization can adapt the model to its own visual identity with the same results.
The benchmark result is promising as a project-reported finding, but it should be read narrowly. The source says Palette ranked first on two of three measured PresentBench domains; the supplied material does not identify the domains, scores, competitors or evaluation setup. That leaves readers without the information needed to judge the size or significance of the lead.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Clicky
Samsung 34-Inch Odyssey G5 Ultra-Wide Curved Gaming Monitor
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Hazelnut Taupe
As an affiliate, we earn on qualifying purchases.
Why Slide Generation Needs Stages
Deck creation combines tasks that are difficult for smaller models to handle reliably in a single generation. The model must decide what information belongs on each slide, then translate that plan into precise positions, sizes and typography without seeing the rendered result as it writes the code. The report describes common failures including overlapping shapes, overflowing text and low-contrast titles.
Palette’s authors also point to a data gap: finished presentations may be available, but they do not necessarily include the decisions or source code that explain how a deck was made. The team used synthetic examples to supply those intermediate steps. Its generated examples include counterfactual rationales—why a chart was chosen over a table, for example, or why a title was reduced in size to fit.
The report says prompt instructions alone did not reliably prevent invented facts when slide specifications asked for content unsupported by the input. In the team’s experiments, providing reference decks changed the teacher model’s output more reliably than adding further rules. That finding underscores a limit of style imitation: matching a visual language and remaining faithful to source content are separate problems.
Benchmark Scope and Transfer Limits
Several details remain unspecified in the supplied report. It does not give PresentBench scores, name the two domains where Palette ranked first, describe the comparison systems or explain how the benchmark evaluation was conducted. The reported ranking therefore cannot by itself show how large the performance gap was or how well Palette performs outside those tests.
The source also does not establish whether the model can reproduce design systems beyond IBM’s, how much human correction generated decks require, or how the workflow performs on real-world confidential material. It describes synthetic training data and a repair harness, but does not report a measured factual-error rate or a general evaluation of content accuracy. The report’s warning that missing information can lead to invented slide content remains relevant.
What Further Evaluation Should Show
The report describes the system’s design and experiments but does not announce a deployment schedule, product release or next benchmark date. A fuller assessment would need published PresentBench scores and methods, alongside results for other design languages and decks based on varied real-world inputs.
For organizations considering this approach, practical tests would include whether the model follows an internal template, preserves source facts, avoids layout defects and reduces editing time after human review. Those results would help determine whether Palette’s staged process works beyond the synthetic examples and IBM-specific training described by its authors.
Key Questions
What is Palette?
Palette is a deck-generation project from IBM Research. It fine-tunes a 20-billion-parameter open-weight model to plan and generate slide presentations in the IBM Design Language.
How does Palette create a presentation?
It first produces a structured design brief for the deck, then generates slide code from each slide’s specification. The report says the brief separates content and layout planning from the coding used to draw slides.
How was the model trained?
The team reports generating about 1,000 decks and 10,000 slides with a large open model. Examples included design rationales and code, and the team rendered and checked slides for problems such as overflow and overlap.
Did Palette outperform other models?
The report says Palette ranked first in two of three PresentBench domains measured by the team. The supplied source material does not include scores, domain names or the full comparison setup, so the result cannot be independently assessed from those details alone.
Can Palette reliably avoid made-up content?
The report does not establish that it can. Its authors say incomplete slide specifications sometimes led the teacher model to invent content, and that reference decks helped more than extra prompt rules in their experiments. They do not provide a measured factual-error rate.
Source: huggingface.co
