Paper2Video
showlab/Paper2VideoReviewed by AI
Checked by RepoPicks Radar on
What is Paper2Video?
Paper2Video is research code from Show Lab (NUS) that makes the presentation video researchers record for a conference. You give it the LaTeX source of your paper, a square photo of the presenter and about ten seconds of their voice; an agent writes the slides, the subtitles and the narration, moves a cursor over the slides, speaks the text in the cloned voice and can add a talking head. There is a lighter mode without the talking head. It runs from the terminal on a machine with a large NVIDIA card and paid AI service keys.
How easy is it to install Paper2Video?
D · terminal, several tools or a specific graphics card · Linux
Linux is the only system with instructions, and it is a long setup: a conda environment with Python 3.10, the tectonic LaTeX engine, API keys exported in the shell, and for the talking head a second conda environment plus cloning the Hallo2 project and downloading its model weights. The README says the minimum recommended card is an NVIDIA A6000 with 48 GB, and its examples pass eight GPUs.
Coming soon: install in 1 click with the RepoPicks app.
Paper2Video on GitHub
- Stars
- 2,422on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +7stars in 6 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Python, NVIDIA graphics card, Other tools
- Disk space
- not determined
- Accounts or keys
- OpenAI, Google Gemini
- Cost
- Free, API costs extra · The code is MIT, but the pipeline calls paid services: the README recommends GPT-4.1 or Gemini 2.5 Pro for both the text and the vision parts. It says locally run open models such as Qwen are also supported, pointing to another project for the details.
- License
- MIT (Permissive)
- Last change on GitHub
- Mar 5, 2026
Paper2Video risk signals (6)
- Attentionclones a face and a voice
README, Inference: the pipeline takes `--ref_img` ("must be square portrait") and `--ref_audio` ("recommended: ~10s") and produces speech in that voice with a talking-head video of that face, using the Hallo2 model. The same steps work with a photo and a voice sample of someone who did not agree to it.
AI review
- Attentionneeds a data center graphics card
README, Inference: "The minimum recommended GPU for running this pipeline is NVIDIA A6000 with 48G", and both example commands pass `--gpu_list [0,1,2,3,4,5,6,7]`, that is eight cards, for cursor generation and talking-head rendering.
AI review
- Attentioncost per video unknown
README, Configure LLMs: the whole paper goes through GPT-4.1 or Gemini 2.5 Pro for text and vision, and slide layout uses tree search (`--if_tree_search` defaults to True), which repeats model calls. The repository gives no estimate of what one video costs.
AI review
- Infopulls in another project and model weights
README, Requirements: for the human presenter you "git clone https://github.com/fudan-generative-vision/hallo2.git", build a separate conda environment for it and download its model weights, because of "potential package conflicts". None of that code is in this repository.
AI review
- Infonot updated for months
The last commit was 2026-03-05, 250 days before this check, there are no releases and the newest README update note is from 2025-10-15.
AI review
- Infobenchmark videos taken from public sites
README, Acknowledgements: "The souces of the presentation videos are SlideLive and YouTuBe", so the dataset it compares against is other people's recorded talks.
AI review
Evidence is shown in its original language.
What Radar could not check about Paper2Video
- how much the OpenAI and Gemini calls cost for one paper, and how often a run has to be repeated
- whether the lighter pipeline without the talking head works on a smaller graphics card
- how much disk the Hallo2 weights and the rendered video need
- whether the slides and narration describe the paper correctly, or invent content a reader would have to check
- whether it runs on Windows or macOS at all, since only Linux-style instructions are given