therealguangzhaohe/PersistBenchDataset

PersistBench: Can 4D Foundation Models Remember?

NeurIPS 2026 Evaluations and Datasets (Spotlight)

Guangzhao He, Hadar Averbuch-Elor*, Wei-Chiu Ma* ("*" denotes equal advising)

Overview

PersistBench is a dataset and evaluation suite for benchmarking visual memory in 4D foundation models, including camera-controlled video generators, video-to-360° video generators, and 4D reconstruction models. It uses 360° videos as reference observations to evaluate objects after they leave the input camera's view, measuring object permanence, motion continuity, and appearance preservation. It contains 2,000 evaluation cases: 1,000 dynamic and 1,000 static.

Run download_persistbench.py to download the source videos and reconstruct the evaluation set locally. The Hugging Face viewer shows metadata and first-frame mask previews; the script creates the RGB frames, masks, cameras, and visibility files used for evaluation.

Standalone dataset download: The instructions below are only for downloading and reconstructing the standalone dataset. To evaluate your model on PersistBench, follow the data setup instructions in our GitHub repository.

1. Set up

Use Linux or macOS, Python 3.11+, and either Node.js 22+ or Deno 2.3+. The JavaScript runtime must be on your PATH for YouTube downloading.

Create an environment and download the repository:

mkdir -p ~/datasets
cd ~/datasets
python3 -m venv .persistbench-venv
source .persistbench-venv/bin/activate
python -m pip install -U pip huggingface_hub

hf download therealguangzhaohe/PersistBenchDataset \
  --repo-type dataset --local-dir PersistBenchDataset \
  --exclude "supplementary/*"

cd PersistBenchDataset
python -m pip install -r requirements.txt

Large supplementary video bundles are fetched automatically when needed. If you already cloned the repository, enter the clone and install requirements.txt in your Python environment.

2. Download and reconstruct

Choose one of the two download options below. Both reconstruct the evaluation set on your local machine. Run these commands from the repository root.

Option 1: Local YouTube download (default)

Download source videos directly on your machine using yt-dlp:

# Test with five cases.
python download_persistbench.py --limit 5

# Download and reconstruct all 2,000 cases.
python download_persistbench.py

Option 2: Google Cloud download

Install the Google Cloud CLI, then choose this option to download sources on a Cloud VM:

# Test with five cases.
python download_persistbench.py --backend cloud --limit 5

# Download and reconstruct all 2,000 cases.
python download_persistbench.py --backend cloud

Follow the script's Google sign-in and project-selection prompts; you do not need to supply a project ID. Your account needs permission to use Compute Engine. The script sets up a worker VM, downloads sources there, transfers them to your local cache, and reconstructs the dataset locally using the same recipes as Option 1.

The VM stays running by default. Add --stop-vm-when-done to stop it after source downloads finish.

Output and cache paths

For either option, reconstructed cases go to ./reconstructed/ and source videos are cached in ./cache/youtube/. A full run reuses completed work from the five-case test. Progress and ETA appear in the terminal.

The --output and --cache flags work with either backend. For example, with local YouTube downloading:

python download_persistbench.py \
  --output /path/to/persistbench_data \
  --cache /path/to/source-cache

Use --limit N for a smaller run or --subset dynamic / --subset static for one subset. Run python download_persistbench.py --help for all options.

Use the reconstructed dataset

The full run creates this evaluation layout:

reconstructed/
├── data.db
├── cases.jsonl
├── .state/                         # Resume checkpoints
├── dynamic/<case_id>/
│   ├── view_0/                      # Input view
│   │   ├── frames/frame_0000.png
│   │   ├── masks/mask_0000.png
│   │   └── camera_params.npy
│   ├── center/                      # Target view; same files
│   └── visibility_dict.npy
└── static/<case_id>/                # Same case layout

Use ./reconstructed/ (or your --output path) as the evaluation dataset root. For model inference and scoring, follow the PersistBench evaluation guide.

Citation

If you use PersistBench in your research, please cite:

@misc{he2026persistbench,
  title={Can {4D} Foundation Models Remember?},
  author={He, Guangzhao and Averbuch-Elor, Hadar and Ma, Wei-Chiu},
  year={2026},
  eprint={2609.20819},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2609.20819}
}