🔍 Read the full analysis: How Palette Uses AI To Create Decks That Match Your Design Style 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 Palette, a 20-billion-parameter open-weight model, to build slide decks in IBM Design Language using separate planning and coding roles. The company reports first-place results in two of three PresentBench domains and a six-to-sevenfold speedup for slide coding, but has not supplied scores, evaluation details or release information.
Hugging Face says it fine-tuned Palette, an open-weight model with 20 billion parameters, to generate slide decks that follow IBM Design Language. The project separates deck planning from slide-code generation, as described in the original analysis, and the company reports that Palette ranked first in two of three PresentBench domains it measured; the available account does not provide scores or enough evaluation detail to independently compare the result.
Palette uses a single LoRA adapter for two roles. Its “designer” role converts a request into a structured JSON brief covering the deck’s content, colors, typography, layouts and design rationale. Its “coder” role then turns each slide’s specification into pptxgenjs JavaScript that defines the slide layout. The project post says the brief can be reviewed or edited before code is written.
Hugging Face says slide coding can run in parallel under this arrangement, making that stage about six to seven times faster than sequential coding. This is the company’s reported speed comparison; the supplied material does not describe the test setup. The team also says coder examples were about five times longer than designer examples, so it oversampled designer examples fivefold during training.
Because the post says there is no public dataset pairing finished decks with their plans, design choices and code, Hugging Face used a larger open model to synthesize training material. The company reports generating about 1,000 decks, or roughly 10,000 slides, across different genres, lengths, layouts and input styles. It rendered the generated decks to check for overlaps and text overflow, repaired broken slides, and added the repairs and related reasoning to the training examples.
A Structured Route to Branded Slides
Palette addresses two practical concerns with automated presentations: whether generated slides follow an organization’s visual identity and where confidential input is processed. Hugging Face presents an open-weight model that organizations could run on their own infrastructure as a possible way to gain more control over both brand consistency and data handling.
Those are project motivations, not demonstrated deployment outcomes. The post does not report organizations using Palette in production or show that running it locally, by itself, protects sensitive information. The two-role design is also relevant beyond this model: separating editorial decisions from layout code makes the deck plan easier to inspect and revise before slides are rendered.
The benchmark claim offers a limited signal that a smaller model can perform well on some measured presentation tasks. Without domain names, scores, competing systems or test conditions, however, the reported rankings do not establish broad superiority over other models or reliability across real-world deck work.
Products Worth Considering
Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Why Palette Splits Planning and Code
A presentation generator must do more than produce text. It has to choose what belongs on each slide and then express that choice through positions, font sizes and visual elements. Hugging Face says models that generate slide code without accounting for the rendered result can produce overlaps, overflowing text or hard-to-read titles.
Palette’s workflow addresses that challenge by producing a deck-wide brief first, then generating the code for individual slides from their specifications. The post also describes training examples that explain why a design choice was made over an alternative. When the team found that incomplete specifications could lead the model to invent missing content, it says examples from reference decks guided the teacher model more reliably than adding further prompt rules.
The project used LoRA fine-tuning, which Hugging Face says made experiments less expensive while adding the deck-building format and skill to the base model. The account focuses on IBM Design Language and does not provide results showing how the method works with other design systems.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s two-role workflow
Benchmark and Release Details Missing
The supplied account does not name the PresentBench domains in which Palette ranked first, provide scores or identify the comparison models and evaluation conditions. The result therefore cannot be checked or compared quantitatively from the information provided. Hugging Face also does not identify the 20B base model in the supplied material.
The post does not state whether model weights, code or evaluation data are publicly available, or give a release date or license. Independent tests are also absent. It remains unclear how Palette handles complex data, strict factual requirements, longer decks or design systems beyond IBM’s, and how often its rendering checks catch layout problems. The account acknowledges that the model may fill gaps in a specification with invented content; it does not quantify how often that occurs.
Likewise, running an open-weight model on an organization’s own infrastructure could offer more control over where inputs are processed, but the post gives no operational safeguards or deployment evidence. Its privacy rationale should not be read as a demonstrated security result.
Evidence Needed for Wider Adoption
The next useful milestones would be details that allow developers and organizations to reproduce the work: a model release and license, the evaluation protocol, PresentBench scores and the identities of the systems used for comparison. The supplied post does not say when any of those details will be available.
Further testing would need to cover different visual systems, deck lengths and tasks that place heavier demands on factual accuracy. Results on the frequency of unsupported content and layout failures—and on how well the checking process detects and repairs them—would help show whether Palette is ready for practical use. Until then, its rankings and speed improvement remain claims reported by the project’s authors.
Key Questions
What is Palette?
Palette is a 20B open-weight model that Hugging Face says it fine-tuned to generate presentation decks in IBM Design Language.
How does Palette create a slide deck?
It uses a designer role to create a structured JSON plan, followed by a coder role that writes pptxgenjs JavaScript for the slides. Hugging Face says the coding stage can run in parallel.
How did Palette perform on PresentBench?
Hugging Face reports that Palette ranked first in two of three measured domains. The supplied account gives no domain names, scores, comparison systems or evaluation conditions, so the result cannot be independently assessed from those details.
Is Palette publicly available?
The supplied project account does not identify a public release date, model license, or availability for the weights and code.
Does Palette protect confidential information?
Hugging Face presents local use of an open-weight model as a potential way to control where inputs are processed. The account does not provide deployment results or security evidence showing how confidential information is protected in practice.
Primary source: Hugging Face · via ThorstenMeyerAI.com
