Can a 20‑Billion‑Parameter Model Build Company‑Branded Slide Decks Without Breaking the Layout?
THE SHIFT
When you ask a big‑language model to draft a presentation, you often get a generic look that drifts toward the model’s own house style. Companies that need decks in a strict brand template – or that cannot let confidential content leave their firewalls – have been forced to either hand‑craft slides or pay for costly proprietary services. The tension is clear: how can you keep full ownership of both data and design while still automating deck creation?
Palette flips that script by offering a small, open‑weight model that you can run on your own infrastructure and that learns your specific design language. Instead of a one‑shot generation, the system splits the task into a designer step that plans the story and a coder step that writes the slide‑drawing code, all while staying inside your security perimeter.
THE PROOF
IBM Research fine‑tuned a 20‑billion‑parameter model on synthetic decks that follow the IBM Design Language. On the public PresentBench benchmark the model topped two of three domains, beating many larger closed systems. The approach also introduced a LoRA adapter that teaches the model the “designer” and “coder” roles without overwriting its base knowledge, keeping training cheap and fast.
To overcome the notorious layout errors of small models – titles disappearing, text spilling off the slide, overlapping shapes – Palette runs every generated slide through a harness. The harness renders each slide, flags geometry problems, and rewrites the offending code only when the fix improves the slide. Across 131 test decks the harness added an average of 3.5 points to the overall score, proving that automated repair can close the quality gap.
THE CATCH
The system’s strength comes with trade‑offs. Training the synthetic deck corpus required a large teacher model to generate about 1,000 decks (≈10,000 slides) and the accompanying reasoning. While the data is fully controllable, it does not capture the full richness of real‑world design decisions, so the model can still hallucinate facts that were never in the brief. The current harness can fix geometry but cannot guarantee factual fidelity without additional reference checks.
Another limitation is that Palette currently relies on a single 20B model; scaling to larger, more diverse corpora may be needed for domains with heavy figure usage or highly creative storytelling. The harness also adds latency – each slide is rendered, checked, and possibly rewritten – which can lengthen generation time compared with a raw model output.
YOUR MOVE
If you manage a design‑centric organization and need to keep presentation data on‑prem, start by evaluating whether your existing slide templates can be expressed in the IBM Design Language format. You can then experiment with Palette’s open‑weight checkpoint (available on Hugging Face) on a small test set, using the provided LoRA adapter to fine‑tune on a handful of your own decks.
While testing, leverage the harness scripts to automatically catch geometry errors and compare the repaired decks against your current manual output. This will give you a concrete sense of quality gains versus the extra compute cost, and will help you decide whether to adopt Palette as a private, brand‑compliant deck generator.
Palette proves that a modest, self‑hosted model can produce on‑brand slide decks at near‑state‑of‑the‑art quality, provided you accept the extra engineering needed for data synthesis and post‑generation repair.
Sources & further reading
Published by The Breakdown. How we report · Our source desk.
