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

The Core Result

PosterMELD achieves an 81.3% success rate at generating print-ready academic posters—posters that actually pass geometric validity, readability, and factual accuracy checks. That's 3.4 times better than the prior best system (P2P) and 5.2 times better than PosterGen. The system also delivers editable PowerPoint files and PNG outputs while costing just $0.38 per poster, a fraction of prior coding-based approaches.

How It Works

The pipeline combines template guidance with multi-agent repair logic. Start with a scientific paper. A "capacity-aware slot" system breaks down what information goes where—title, key figures, methodology, results—before any rendering happens. This writing-first, rendering-second order prevents the common failure mode where an image generator tries to fit too much text or misaligns elements.

When writing is complete, the system renders to a visual canvas. Here's where the repair loop kicks in: deterministic checks (does text fit? are fonts readable?) catch obvious failures immediately. If something fails, a vision-language model (VLM)—a neural system that understands both images and text—reviews the problem and decides whether it's worth fixing. Not every failure is repairable within budget or time; the VLM acts as a gatekeeper that routes only fixable issues to bounded repair attempts.

The output is native: users get both an editable PowerPoint file and a print-ready PNG. This dual format is harder than it sounds. Most systems either generate images (not editable) or generate code (slow and expensive). PosterMELD maintains both formats through its template structure—the same slot assignments that created the poster can be exported as editable objects.

Explicit design controls let you ask for variations of the same poster (different color schemes, different layout styles) without regenerating from scratch. This controllability is retained despite the low cost because the template abstraction doesn't require calling expensive APIs repeatedly.

Why It Matters

Academic poster generation is a real pain point. Papers are dense; posters must distill them into something readable in 30 seconds. Prior systems either hid failures (only reported scores on outputs that didn't crash), required expensive custom code per poster, or produced outputs locked in image format—unable to tweak a typo or adjust a figure without regenerating everything.

PosterMELD solves three problems simultaneously: reliability (the gated repair loop and VLM review push the success rate into the 80s), editability (PowerPoint export and explicit design controls), and cost (0.38 per poster). The evaluation across 621 papers uses a strict Print-Ready Rate (PRR) metric that counts only posters passing geometry, readability, asset integrity, and factual sanity checks—not soft scoring on completed samples. A frozen VLM then scores the passing outputs on craftsmanship, harmony, and expressiveness, confirming that PosterMELD's pass-through outputs rival or exceed generated alternatives.

The system won't replace human designers for high-stakes conference posters, but it moves the needle significantly for batch processing and rapid iteration, where editability and reliability matter more than photographic perfection.