Overview

PersistBench is a dataset and metric suite for evaluating visual memory in 4D foundation models, including camera-controlled video generators and 4D reconstruction models. It uses 360° videos as reference observations to assess objects after they leave the input camera's view, measuring object permanence, motion continuity, and appearance preservation. See the paper abstract.

This early code release supports your own model and the reference diagnostic.

Release checklist

• Custom model inference and evaluation
• Dataset download workflow
• PyPI publication
• Release code for baselines in the paper

Setup

Use Python 3.10+ in your model environment. Replace /path/to/... and your-model-env with your own paths and Conda environment name.

1. Clone PersistBench into your workspace:

   git clone https://github.com/guangzhaohe/PersistBench.git /path/to/PersistBench
   

2. Activate your model's existing Conda environment, then install from the PersistBench repo:

   conda activate your-model-env
   cd /path/to/PersistBench
   python -m pip install -e .
   

   Alternatively, install from PyPI in your model environment: python -m pip install persistbench.

PyPI only (0.1.1+): run persistbench docs for the complete workflow, or persistbench docs --output /path/to/persistbench-workspace to export the guides, model example, and Qwen launcher. Follow the PyPI guide without cloning this repo.

Dataset

Prepare the dataset separately at /data/benchmark. The download workflow is coming; currently use the dataset format.

Inference

3. Go to your model's repo, keeping your model environment active:

   cd /path/to/your-model
   

4. Create my_model.py in that repo. Replace the loading and generation calls below with your model's API; set its frame count and resolution:

   from persistbench import BaseModel, ModelOutput
   
   class MyModel(BaseModel):
       num_frames = 49
       resolution = (384, 512)  # height, width; -1 preserves dataset shape
   
       def __init__(self, device="cuda:0"):
           super().__init__(device=device)
           self.model = load_your_model(device=device)
   
       def predict(self, sample):
           frames = self.model.generate(sample.input_frames,
                                        sample.input_cameras, sample.target_cameras)
           return ModelOutput(frames)  # CPU NumPy RGB [T, H, W, 3]
   

   See the executable example and interface details.
5. Run inference from your model's repo, in the same model environment:

   persistbench infer --model ./my_model.py:MyModel --name my-model --dataset /data/benchmark
   

   Predictions are saved to /path/to/your-model/outputs/null/my-model/.

Evaluation

1. Create the evaluation Conda environment. Run from the PersistBench repo:

   cd /path/to/PersistBench
   conda create -n persistbench-eval python=3.11 pip -y
   conda activate persistbench-eval
   python -m pip install '.[eval]'
   

2. Install the metric models and download their checkpoints, in this same environment and repo: follow the complete checkpoint guide (SAM2, DINOv3, DINOv2).
3. Start or connect to Qwen. Follow the independent Qwen judge guide. Hosting uses a separate Conda environment and terminal; keep the evaluation terminal open.
4. Evaluate and aggregate. In the evaluation terminal, from the PersistBench repo:

   cd /path/to/PersistBench
   conda activate persistbench-eval
   export SAM2_CHECKPOINT="$PWD/checkpoints/sam2.1_hiera_large.pt"
   export DINOV3_REPO="$PWD/vendor/dinov3"
   export DINOV3_CHECKPOINT="$PWD/checkpoints/dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth"
   export PERSISTBENCH_VLM_BASE_URL=http://127.0.0.1:8000/v1
   persistbench evaluate --dataset /data/benchmark --predictions /path/to/your-model/outputs/null/my-model
   persistbench aggregate --results /path/to/your-model/outputs/persist_bench/my-model
   

   For a remote Qwen server, replace the endpoint above with its URL.

Read /path/to/your-model/outputs/persist_bench/my-model/summary.json for static/dynamic × visible/invisible averages and valid-case counts. Evaluation saves final scores; aggregation only averages them. See score definitions for formulas and missing-score handling. Commands use the active Conda environment; they do not switch environments.

Visualizations and progress

Per-case evaluation folders include matching images, tracked-mask overlays, object crops, and judge inputs. From the PersistBench repo, in the evaluation environment:

persistbench dashboard --output /path/to/your-model/outputs --port 8080

Open http://127.0.0.1:8080. For splitting, resuming, and configuration, see usage; for checks and limitations, see validation.

Acknowledgements

We thank the authors of SAM2, DINOv3, DINOv2, and Qwen for their open-source projects.

License

This project is licensed under the MIT License.

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}
}