Infographics stop being repainted whole: text stays vector, illustrations generate in layers, and rewording no longer needs a diffusion pass
Synopsis
Infographic generation for AI learning assistants now splits the job: text and layout stay in an editable vector layer while non-text illustrations come from parallel layer-set diffusion, keeping typography accurate while scoring higher on overall aesthetic quality.
Interpretation
Infographic generation is split into two jobs: a deterministic vector structure carries geometry, hierarchy, occlusion order and text, while diffusion only makes the non-text content look rich. Each prior path supplied half: vector representations preserved geometry, hierarchy and text but looked rigid, while raster diffusion models produced richer appearance by flattening content and layout into one image. On shared held-out layouts the hybrid representation beats both raster baselines on every text metric and approaches the vector reference, which places exact text by construction.
The method partitions vector elements into ordered, mutually non-overlapping layer groups, processes each group through a parallel RGB generation stream, and composites them in paint order. Rasterizing the vector first and conditioning on it preserves coarse placement but discards element ownership, occlusion relationships and the structure needed to recompose the graphic. Training records were synthesized by a multi-agent pipeline into 23,960 paired infographic samples, each carrying a structural vector, a text-free target, per-layer amodal and visible masks, paint order, and a hierarchy and overlap graph.
Text nodes stay outside the diffusion process entirely and are rendered directly from preserved vector metadata, so changing wording, typography or language needs no further diffusion pass. End-to-end raster generators must repaint text whenever wording, typography or language changes, and their data-driven glyph generation may favor the conventions of a dominant training language. The same generated layers are shown with text rendered in four languages, where only the text node attributes are replaced and the image layers stay unchanged.
Perspective
The representation targets authoring workflows where instructional graphics must be revised and reused: text, layout and layer identity remain addressable after generation, so changing copy, switching language, moving or hiding layers, reusing illustrations, specifying animation trajectories, and reflowing the same content onto slide, poster and mobile layouts all change structural metadata without regenerating image layers. It requires a valid source vector structure as input for those properties to hold.
The method requires a valid source vector and inherits errors and biases in its content, layout and training data. Deterministic typography preserves supplied text but does not ensure factuality, accessibility or culturally appropriate localization, and edits that change occlusion or z-order may affect unintended regions. The authors also note that generalizing to more layouts and improving transformations under changed occlusion remain open research directions.
