IBM Research has described Palette’s method for training a model to plan presentation design and write slide code through one shared adapter. The system creates an editable design specification before drawing pages, then checks the rendered slides for layout errors. The team’s published account describes its training examples, reference revisions and separate checking software.
IBM Research has described Palette, an experimental approach to producing presentation slides with a model trained in IBM’s visual design language. This account comes from the research team’s own November publication. A shared adaptation layer gives the model two jobs: deciding the deck’s overall design and writing the instructions that draw its individual pages. The design stage produces an editable specification before slide generation begins.[1]
The designer chooses such elements as colours, typography, content and page arrangement. The coding role then handles a single slide at a time, using the JavaScript slide-generation library pptxgenjs. Separate pages can be processed concurrently because their design has already been settled. Training examples generated with a teacher model include explanations of design choices and rejected alternatives. The team increased the share of designer examples to balance the longer coding examples.[1]
The surrounding software renders the pages and looks for overlapping, clipped or overflowing elements. A separate editing model revises affected slide code, and the system retains a revision only when its geometric checks improve. During development, the team changed reference examples after finding invented table material in outputs. For evaluation, a different planning model turned source documents into instructions for Palette. The published work describes the training and checking method.[1]
