WorldMemBench evaluation weights
Model packages are stored once under models/, retaining their upstream names. The WorldMemBench downloader creates relative symlinks under metrics/<metric>/ after downloading. Upload model files, not the generated metrics/ symlinks.
For upload, exclude metrics/**, **/.cache/**, **/blobs/**, and *.zip. ViPE's Hugging Face cache snapshot files must be uploaded as their actual contents; the Hub package needs their existing models/vipe/huggingface/hub/.../snapshots/... paths, not duplicate blob-cache files.
python scripts/download_release.py weights --datasets-dir /path/to/datasets
# Optional Dynamic Memory model
python scripts/download_release.py weights --metrics dynamic_memory --datasets-dir /path/to/datasets
export WORLDMEMBENCH_DATASETS_DIR=/path/to/datasets
The default downloads all non-Dynamic metric weights. Use --metrics imaging_quality or another metric name to download only the corresponding models.
Every model retains its upstream license and model-card terms; the WorldMemBench code's MIT license does not relicense these checkpoints. This package is a staging directory: verify redistribution rights for each model before publishing files. In particular, InsightFace models have non-commercial research terms and the selected DA3 checkpoint has CC BY-NC 4.0 terms.
The code repository retains scripts/download_weights.py for explicit downloads from the original upstream sources, including the opt-in Dynamic Memory model.