rescuerz/WorldMemBench-weights

WorldMemBench Weights

Pretrained model weights for WorldMemBench, a benchmark for memory in video world models.

This repository gathers the models used for video quality, Human Memory, Object Memory, Environment Memory, Dynamic Memory, and Camera Following evaluation. Shared models are stored once under models/.

Usage

Follow the WorldMemBench installation instructions to set up the evaluation environment. From the WorldMemBench code directory, download this model repository:

hf download rescuerz/WorldMemBench-weights --local-dir ./weights

# Check local SHA256 values and create the per-metric model links
python scripts/download_weights.py --verify

The full download includes Qwen3.8-27B for locally hosted Dynamic Memory evaluation. For a smaller, metric-specific download from the original upstream sources, use the code repository's downloader instead:

python scripts/download_weights.py --metrics imaging_quality aesthetic_quality

To use another directory, pass --local-dir /path/to/weights to hf download, then --weights-dir /path/to/weights to the verification and evaluation commands. Downloading weights does not start the Dynamic Memory VLM server or install the DA3/ViPE environments; see the code repository for these steps.

Included models

Evaluation	Models
Imaging / Aesthetic Quality	MUSIQ-SPAQ, CLIP ViT-L/14, LAION Aesthetic Predictor
Human Identity / Appearance	InsightFace buffalo_l, YOLO-World, CLIP ViT-B/32, SAM2; DINOv2-Large for Appearance
Object Appearance / Geometry	Grounding DINO Base, CLIP ViT-B/32, SAM2; DINOv2-Large for Appearance
Environment Memory	DA3NESTED-GIANT-LARGE-1.1, SALAD; DINOv2-Large for Visible Memory; SegFormer for Scene Reconstruction
Dynamic Memory	Qwen3.8-27B
Camera Following	DA3 and SALAD; ViPE's SAM, DeAOT, DROID-SLAM, GeoCalib, Grounding DINO, BERT and UniDepthV2 dependencies for Translation

Directory layout

weights/
├── README.md
├── LICENSE_NOTICE.md
├── models/                  # model files and their accompanying configuration
│   ├── buffalo_l/
│   ├── sam2-hiera-large/
│   ├── dinov2-large/
│   ├── Qwen3.8-27B/
│   ├── vipe/
│   └── ...
└── metrics/                 # relative symlinks created locally by the script
    ├── human_identity/
    ├── human_appearance/
    └── ...

Keep the nested paths inside models/vipe/: its loaders use the supplied Hugging Face snapshot files and revision references. The generated metrics/ links are not separate copies of the weights.

Licenses and disclaimer

Each model retains its own upstream license and usage restrictions. This collection is not covered by a single MIT or Apache-2.0 license. The WorldMemBench code license does not relicense these weights. See LICENSE_NOTICE.md for the model sources and licensing notes, including non-commercial restrictions.

This collection is intended to support research and benchmark evaluation. That purpose does not override any upstream condition or grant additional redistribution rights. For licensing or attribution concerns, please open an issue in WorldMemBench with the affected model and supporting details.

Acknowledgements

We thank the original model authors and maintainers for making their work available. Please also cite the relevant upstream papers when using their models in your research.