IBM Research released Palette, a methodology to fine-tune a 20B open-weight model for generating slide decks. It uses a LoRA adapter to separate editorial planning from geometric coding, allowing parallel slide generation. The system is trained on approximately 10,000 synthetic slides using the IBM Design Language.
This is useful if you need self-hosted, brand-compliant deck generation without sending confidential data to frontier APIs. The catch is the heavy reliance on synthetic data and a specific design language, which limits immediate generalization. You must be comfortable building a verification harness to catch layout errors in the generated code.
more: https://huggingface.co/blog/ibm-research/palette-deck-building
my other side: ivo.blogspot.com
Leave a Reply