From PDF to Presentation: Turning Academic Papers Into Narrated Videos

You've probably sat through enough dry academic talks to know the pain of turning a dense paper into a clear presentation. It's a tedious process—reading, summarizing, designing slides, recording narration, syncing everything together. What if the whole pipeline could be automated? That's exactly what Paper2Video aims to do: take a scientific paper, an image, and an audio clip as input, and produce a complete presentation video as output.

The project comes from Show Lab at the National University of Singapore, and it's built around two core components: a generation system called PaperTalker and an evaluation benchmark that shares the project's name. It was accepted to the Scaling Environments for Agents Workshop at NeurIPS 2025, so this isn't just a weekend hack—it's a serious research contribution.

What It Does

Paper2Video is really two things working together.

On the generation side, PaperTalker is an agent that takes a paper, a portrait image (like the author's photo), and a reference audio sample, then produces a narrated presentation video. It handles the whole pipeline internally: creating slides, adding subtitles, grounding cursor movements to what's being discussed, synthesizing speech, and rendering a talking-head video that matches the provided image and voice.

On the evaluation side, the Paper2Video benchmark provides metrics to assess how good these generated presentation videos actually are. That's an important piece—if you're going to automate video generation, you need a way to measure quality beyond just eyeballing the output.

The architecture is modular, which means you can swap components in and out. You'll need to configure LLMs for the slide generation and scriptwriting parts, but the system handles the rest of the pipeline from there.

Why It's Cool

What makes this project stand out is how it tackles the full problem rather than just a piece of it.

• It's a complete pipeline, not a single trick. Many video generation projects focus on one aspect—maybe lip-syncing or slide creation. Paper2Video integrates slides, subtitles, cursor grounding, speech synthesis, and talking-head rendering into one agent. That's a lot of moving parts working together.

• It works with what you give it. The input is refreshingly simple: a paper, an image, and an audio sample. You don't need a studio recording of the author's voice or professionally designed slides. The system handles the transformation from those raw inputs to a finished video.

• It includes an evaluation benchmark. This is rare and valuable. Most projects just show you pretty demos. Paper2Video ships with metrics designed specifically to evaluate presentation quality, which means you can actually compare approaches and measure improvement.

• There's a fast mode. The latest update includes a version without the talking-head component for faster generation. That's a practical touch—sometimes you want the full avatar experience, sometimes you just need a quick narrated slide deck.

• The demo is wild. The project page shows an example using Geoffrey Hinton's famous dropout paper, his photo, and a reference audio clip. The result is a presentation that looks like Hinton himself is presenting it. It's genuinely impressive to see a paper you've read turned into a video presentation automatically.

How to Try It

If you want to try it on your own paper, the setup is straightforward. The code lives in the src directory, and you'll need Python 3.10 and a conda environment:

cd src
conda create -n p2v python=3.10
conda activate p2v
pip install -r requirements.txt
conda install -c conda-forge tect

After that, you'll need to configure your LLM access—the system relies on LLMs for the script and slide generation parts. Then you can run inference with your paper, image, and audio as inputs.

Check out the repository for the full instructions and examples. There's also a dataset available on Hugging Face and a project page with more sample videos if you want to see what the output looks like before you commit to setting it up.

Final Thoughts

Paper2Video is best suited for researchers who regularly create presentation content, educators who want to repurpose papers into lecture videos, or anyone curious about where automated content generation is heading. It's not going to replace your presentation skills—but it might save you hours on the mechanical parts of video creation.

The fact that the team built an evaluation benchmark alongside the generator shows they're thinking about this as a research problem, not just a demo. That's the kind of rigor that pushes a field forward. If you've ever wanted to turn your latest paper into a narrated video without spending a weekend on slides and recording, this is worth a look. The project is open source and welcomes contributions, so if you spot a weak spot in the pipeline, you can jump in and improve it.