PosterMELD: AI Pipeline Generates Editable Scientific Posters from Papers

A new multi-agent system named PosterMELD has been developed by researchers to automatically create editable, print-ready scientific posters from lengthy multimodal documents. This innovative approach overcomes the shortcomings of previous methods by ensuring that each element can be edited and effectively managing any errors during the creation process. Utilizing a template-conditioned framework, PosterMELD features capacity-aware slots that guide the writing phase prior to rendering, while deterministic gates and a vision-language model (VLM) handle failures through controlled repair mechanisms. The system can generate both editable PowerPoint (PPTX) and Portable Network Graphics (PNG) files, allowing for various versions of the same paper. Evaluated on 621 papers, it achieved a Print-Ready Rate (PRR) of 81.3%, which assesses compliance with geometric, readability, asset integrity, and factual accuracy criteria. Native editability is reported separately, and a frozen VLM provides Craftsmanship-Harmony-Expressiveness (CHE) scores for the outputs. This advancement holds great promise for enhancing the efficiency of academic poster creation, potentially benefiting researchers and designers alike. The paper can be accessed on arXiv with the identifier 2608.02218.