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

Imagine you are a scientist who has just finished writing a massive, complex research paper. It's full of charts, equations, and dense paragraphs. Now, you need to present this at a conference, but instead of a slideshow, you have to fit all that information onto a single, giant poster board. This is a tricky job. You can't just shrink the paper down; you have to rearrange the story, pick the best pictures, and make sure the text is big enough to read from three feet away. If you get it wrong, the text might spill over the edge, the pictures might get squished, or the whole thing might look messy and unreadable. For decades, making these posters has been a tedious, manual chore, like trying to fit a king-sized bed into a dollhouse without breaking the frame.

Recently, computers have started trying to help with this "paper-to-poster" magic. Some systems just take a picture of the final result, which looks nice but is impossible to fix if you make a typo. Others try to write code to build the poster, which is powerful but often costs a fortune in computer time and can still produce messy results. The big question researchers are asking is: Can we build a system that is smart enough to design a beautiful poster, strict enough to ensure it's actually ready to print without errors, and flexible enough to let you tweak the design later, all without breaking the bank?

Enter PosterMELD, a new system that acts like a team of specialized architects and editors working together to solve this puzzle. Think of it as a construction crew that doesn't just throw bricks at a wall and hope they stick. Instead, they start with a blueprint. Before they write a single word of text or place a single image, they look at a library of proven poster layouts (templates) and assign specific "slots" for content. These slots are like pre-measured containers: "This box holds exactly 50 words," "This space fits one large graph," and "This area is for the title."

The system uses a team of five digital agents (smart computer programs) to do the work. First, a Template Agent picks the right blueprint and sets the rules. Then, a Content Agent writes the text, but it's forced to stay within the size limits of the slots, so nothing ever gets crammed or cut off. A Layout Agent arranges the pieces, and a Visual Agent adds the colors and images, making sure everything is editable—meaning you can still click on a text box to change a word later, just like in a standard PowerPoint file. Finally, a Review Agent acts like a strict inspector. It checks the draft against a list of hard rules: "Is any text too small to read?" "Are any images missing?" "Is the text overlapping?" If it finds a problem, it sends just that specific part back to be fixed, rather than scrapping the whole poster.

The results of this approach are impressive. The researchers tested PosterMELD on 621 different scientific papers. They found that 81.3% of the time, the system produced a poster that was perfect, printable, and editable. To put that in perspective, other similar systems only succeeded about 24.2% or 15.8% of the time. While a system that just generates a single image (like GPT-Image-2) had a slightly higher success rate of 85.2%, its output was a flat picture that couldn't be edited. PosterMELD managed to keep the high success rate while keeping the ability to edit the text and images.

Perhaps the most exciting part is the cost. The system that uses coding agents to build posters (Codex+Skill) costs about $10.78 per poster to run. PosterMELD does the same high-quality job for just $0.38 per request. That's a massive difference, making it possible for almost anyone to generate professional-grade posters. The system also allows for "design diversity," meaning you can ask it to make the same poster look completely different—changing the colors, the layout style, or the density of information—just by flipping a switch, giving you multiple options to choose from.

In short, PosterMELD suggests that by planning the structure first and using a team of specialized, rule-following agents, we can automate the creation of scientific posters in a way that is reliable, cheap, and fully editable. It doesn't just guess; it builds with a plan, checks its work, and fixes only what's broken, ensuring that the final product is ready to be printed and pinned up on a conference wall.