PosterMELD

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

Multi-Agent · Editable · Layouts · Design diversity

Project Page · Poster Generation · Benchmark Evaluation · Quick Start

PosterMELD converts scientific papers into native, editable PowerPoint posters and matching PNG renders. It plans template capacity before writing, grounds every section in the source paper, composes figures and tables as editable assets, and applies bounded deterministic and VLM quality checks.

Repository Modules

The repository is organized as two independently runnable modules:

Module	Purpose	Entry point
poster_generation/	End-to-end PDF-to-PPTX/PNG generation pipeline, including prompts, templates, assets, configuration, scripts, and regression tests	python -m src.workflow.pipeline or postermeld
benchmark_eval/	Standalone PRR/CHE, Universal Score, and Keypoint BERTScore evaluation code	python -m prr_che.evaluate, python -m universal_score.evaluate, and python -m keypoint_bertscore.*

The project website and paper figures are kept in docs/ so GitHub Pages remains independent of both runtime modules.

Highlights

• Multi-agent composition: specialized agents coordinate paper understanding, capacity-aware writing, layout, visual placement, and review through one traceable state.
• Editable outputs: text, figures, tables, section bars, and logos remain native PowerPoint elements.
• Design diversity: 16 landscape and 8 portrait templates support controllable style, density, background, header, and seed variations.
• Paper-grounded content: MinerU extracts text and visual assets; Marker is retained as an automatic fallback.
• Bounded quality repair: deterministic geometry checks and VLM review target only failed aspects instead of repeatedly rebuilding the poster.
• Reproducible evaluation: request-level PRR, conditional CHE, Universal Score, and Keypoint BERTScore are provided as a separate package.

Results at a glance

Method	PRR (%) ↑	CHE ↑	Universal ↑	Editable	Cost / request ↓
GPT-Image-2	85.2	2.698	4.948	No	$0.18
Codex+Skill	82.8	2.716	4.876	Yes	$10.78
Paper2Poster	0.2	-	2.772	Yes	$0.34
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 computed over all requests. CHE is conditional on print-ready outputs, while Universal Score keeps missing generations in the full-benchmark aggregate.

Method

The generator follows a structure-first pipeline:

1. Parse the PDF into grounded text, figures, tables, formulas, and metadata.
2. Select a template and expose slot geometry, reading order, and capacity budgets.
3. Distill poster keypoints and compose sections within those budgets.
4. Place visuals, render an editable PPTX, and produce a consistent PNG preview.
5. Apply local and global quality checks, followed by bounded repair when needed.

Quick Start

Generate a poster

git clone https://github.com/Jackey0903/PosterMELD.git
cd PosterMELD/poster_generation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e . --no-deps
cp .env.example .env

postermeld data/0409_demo/paper.pdf \
  --layout-template auto \
  --disable-generated-teaser \
  --disable-generated-background

Enable the complete visual pipeline after configuring the VLM, image-generation, and MinerU endpoints in .env:

postermeld /path/to/paper.pdf \
  --layout-template auto \
  --poster-style navy_serif \
  --visual-density rich \
  --enable-generated-teaser \
  --enable-generated-background \
  --enable-vlm-layout-review \
  --enable-visual-legibility-review

See the generation guide for all controls, output files, templates, and backend configuration.

Evaluate generated posters

cd ../benchmark_eval
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

python -m prr_che.evaluate \
  --manifest examples/manifest.jsonl \
  --output-dir outputs/prr_che \
  --model gpt-5.5

See the evaluation guide for manifest schemas, aggregation rules, GPU evaluation, and reproducibility details.

Benchmark

PosterMELD was evaluated end-to-end on 621 papers across 14 publication-source groups and 10 research domains. The evaluation module preserves missing generations in request-level denominators and separates print readiness from conditional visual quality.

Metric	Scope
PRR	Binary request-level print readiness
CHE	Craftsmanship, Harmony, and Expressiveness for print-ready outputs
Universal Score	Ten reference-aware poster criteria with missing outputs scored as zero
Keypoint BERTScore	OCR-based content coverage against ordered paper keypoints

Project Layout

PosterMELD/
├── poster_generation/        complete generation subproject
│   ├── assets/               conference and bundled visual assets
│   ├── config/               prompts and pipeline configuration
│   ├── data/                 runnable example papers
│   ├── scripts/              batch-generation utilities
│   ├── src/                  agents, layout, tools, state, and workflow
│   ├── templates/            16 landscape + 8 portrait templates
│   ├── tests/                generation regression tests
│   └── requirements.txt
├── benchmark_eval/           standalone benchmark-evaluation subproject
│   ├── common/               manifest, API, image, and JSON utilities
│   ├── prr_che/              print-ready and conditional aesthetic metrics
│   ├── universal_score/      reference-aware universal evaluation
│   ├── keypoint_bertscore/   OCR and content-coverage evaluation
│   ├── examples/             manifest and annotation examples
│   └── tests/                offline evaluation tests
├── docs/                     GitHub Pages project site and paper figures
├── Makefile                  repository-level validation commands
└── LICENSE

Validation

Run both offline validation suites from the repository root:

make test

Or validate each module independently:

make test-generation
make test-evaluation

License

PosterMELD is released under the MIT License.

One paper, many valid posters.