Multi-agent paper-to-poster generation

PosterMELD Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs

M Multi-Agent E Editable L Layouts D Design diversity

Fix the structure first, then write to fit. PosterMELD turns one paper into multiple valid poster designs while preserving evidence, editability, and print-readiness.

Code Quick start Benchmark

81.3%print-ready rate

3.4×P2P's print-ready rate

621benchmark papers

$0.38mean cost per request

Paper teaser

Editable. Compact. Readable. Budgeted.

PosterMELD combines structure-first composition with bounded quality repair, retaining the practical requirements that other generation routes sacrifice.

Comparison of P2P, PosterGen, a coding-agent workflow, and PosterMELD

01 · Overview

A poster is not a shrunken paper.

Scientific poster construction compresses a long multimodal paper into one readable, editable canvas. Existing systems can hide request-level failures by scoring only completed outputs, while direct image generation does not preserve element-level editability.

PosterMELD inverts the usual workflow. Capacity-aware slots guide writing before rendering; specialized agents select evidence, compose visuals, and export native PowerPoint elements. Deterministic gates and VLM review route only failed aspects to a bounded repair loop.

Print-ready by construction

Capacity-aware writing and native PPTX rendering prevent overflow, overlap, and unreadably small text.

Controlled one-to-many output

Template, style, density, background, and seed controls produce meaningful alternatives instead of accidental variation.

Validity-centered evaluation

A 621-paper benchmark reports request-level Print-Ready Rate alongside conditional CHE, content fidelity, editability, and cost.

02 · Method

Structure first. Reflection after rendering.

Paper understanding, template-conditioned writing, reusable skills, native rendering, and bounded repair operate as one auditable pipeline.

PosterMELD framework from paper inputs and controls through agent-skill composition, bounded VLM-guided repair, and diverse editable outputs — PosterMELD combines four specialized agents, reusable skills, capacity-aware slot contracts, deterministic gates, and VLM review. Failed drafts receive bounded rewrite, reflow, resize, or rerender actions; accepted drafts become editable PPTX files with consistent PNG renders.

Parse and ground

MinerU extracts text, figures, tables, and document structure while provenance remains attached to every asset.

Declare capacity

Template slot contracts expose geometry, reading order, prominence, character budgets, and compatible visual types.

Compose to fit

Content, template, layout, and visual agents select keypoints and place evidence against known spatial constraints.

Review and repair

Deterministic gates and VLM critics inspect grounding, overflow, readability, visual quality, and occupancy before acceptance.

03 · Controlled diversity

One pipeline, distinct visual identities.

The final paper compares five domains. PosterMELD changes layout topology, typography, palette, density, and text-figure allocation while keeping every output natively editable.

PosterMELD outputs from the paper

Medical imaging, binary analysis, virtual dressing, fairness, and population genomics.

PosterMELD poster for coronary angiography translation correction — Medical imaging · clinical blue

PosterMELD poster for BDLF-Qwen3 binary function similarity — Binary analysis · burgundy

PosterMELD poster for AnyDressing virtual garment generation — Virtual dressing · teal

PosterMELD poster for the AI-Face fairness benchmark — Fairness benchmark · gold + navy

PosterMELD poster for a population-scale copy-number variation database — Population genomics · navy + green

Qualitative comparison of P2P, PosterGen, GPT-Image-2, and PosterMELD on five papers — Final qualitative figure from the paper. Each column uses the same input paper; each row shows one generation method.

04 · Template library

Templates declare capacity, not content.

16: landscape topologies
8: portrait topologies
3 × 3: style and density profiles

TemplateStyleDensityBackgroundSeed

PosterMELD template library construction from real posters — Real posters are decomposed into semantic blocks, normalized as spatial features, and clustered into reusable template structures.

05 · Benchmark and evaluation

Validity is an outcome, not an assumption.

The benchmark spans 14 publication-source groups and ten research domains. PosterMELD evaluates whether each request produces a usable poster before reporting conditional visual and content scores.

Conference and journal composition of the 621-paper PosterMELD benchmark

Print-Ready Rate

Request-level validity across geometry, readability, asset integrity, and obvious factual errors.

Content fidelity

Paper-grounded keypoints, evidence coverage, and figure-text consistency.

Conditional CHE

Craftsmanship, Harmony, and Expressiveness are scored only for print-ready outputs.

Request cost

Mean generation cost includes failed requests instead of hiding them from the denominator.

235 NeurIPS 101 AAAI 100 CVPR 92 Journals 93 Other conferences

End-to-end comparison

High validity without giving up editability.

All metrics use the shared 621-paper request set. Missing generations remain in the denominator.

End-to-end paper-to-poster results over 621 requests
Method	PRR ↑	CHE ↑	Universal ↑	Editable	Cost ↓
GPT-Image-2	85.2%	2.698	4.948	No	$0.18
Codex+Skill	82.8%	2.716	4.876	Yes	$10.78
P2P	24.2%	3.071	4.033	No	$0.35
PosterGen	15.8%	3.163	3.901	Yes	$0.28
PosterMELD	81.3%	3.247	4.456	Yes	$0.38
Human reference	98.7%	3.287	4.995	Yes	—

PRR is request-level print readiness. CHE is conditional on print-ready outputs. Universal assigns zero to missing generations. Cost is the mean U.S. dollars per request, including failures.

Open source

Generate an editable poster from your paper.

Start with a standard template, then control style, density, background, and review depth from the command line.

Explore the code Read quick start