PosterMELD Open-Sources an Editable Research Poster Agent Pipeline That Constrains Content Generation by Layout Capacity First

Turning a paper into an academic poster is not simply a matter of summarization: the system must also allocate limited layout space, preserve relationships between figures and text, prevent text overflow, and produce a file that authors can continue editing. Released on August 3, PosterMELD uses a multi-agent architecture that “allocates capacity first, then writes content,” avoiding the need to force fully generated text into a template afterward. The project has open-sourced its generation pipeline, evaluation code, templates, and tests under the MIT License.

The workflow first uses MinerU to parse text, equations, images, tables, and metadata from PDFs, with Marker available as a fallback. The system then selects a layout from 16 landscape and eight portrait templates and passes each section’s geometric dimensions, reading order, and word capacity to the content agent. Subsequent agents extract key points from the paper, place visual assets, and generate a PPTX file. Text, charts, title bars, and logos remain native PowerPoint elements, while a PNG preview is rendered separately.

Quality control does not allow agents to revise their work indefinitely. Deterministic gates first check for overlaps, out-of-bounds elements, readability, and asset integrity, after which a vision-language model reviews the overall layout. Only failed items enter a bounded local-repair loop. The team therefore defines a request-level Print-Ready Rate (PRR), counting requests that fail to produce an output as failures, while the visual-quality CHE score is calculated only for outputs that have passed the print-readiness checks.

In the team’s custom evaluation covering 621 papers across 10 research fields, PosterMELD achieved a PRR of 81.3%, compared with 24.2% for P2P and 15.8% for PosterGen. Its average cost was $0.38 per request. GPT-Image-2 achieved a higher PRR of 85.2%, but its outputs could not be edited element by element. PosterMELD’s main value, therefore, lies in combining generation quality, editability, and cost within a single executable pipeline rather than leading on any individual visual-quality score.

The reported figures were produced by the authors using their own dataset, a frozen VLM judge, and a specific combination of APIs. Because CHE compares only successful outputs, it cannot independently represent quality across all requests. Deployers must also verify that OCR results, equations, and citations remain faithful to the source material, and remeasure costs and reproducibility when using different model providers.