archives

Original compressed Kubric-4D dataset files, mirrored from the Generative Camera Dolly project. Credit for the dataset belongs to its original creators. This repository does not change the original data or its usage terms.

Access to data files requires a Hugging Face account and manual approval from the repository owner. The original collection contains 20 archive files totaling 7,652,279,103,561 bytes. Uploading is in progress. Large originals are stored as binary .partNNN pieces (approximately 10–16 GB each) to allow reliable verification. Every original listed in manifests/ has all of its pieces available and verified. Other originals may still be pending.

After downloading the repository, restore the original compressed files with:

python3 restore_archives.py

The script verifies every piece and the restored original using SHA-256, and refuses to overwrite existing originals. It does not extract archives. See RESTORE.md for download commands, disk-space requirements, and restoring a single original. The table below describes the original files after restoration.

Files	Count	Bytes per file
gcd_kubric4d_train.tar.gz.aa through .aq	17	400,000,000,000
gcd_kubric4d_train.tar.gz.ar	1	341,393,715,036
gcd_kubric4d_val.tar.gz	1	256,464,456,525
gcd_kubric4d_test.tar.gz	1	254,420,932,000

After restoring their .partNNN pieces, the 18 training files are consecutive pieces of one gzip-compressed tar stream. They are not independently extractable. After downloading all files, they can be joined in alphabetical order, for example:

cat gcd_kubric4d_train.tar.gz.a? | tar -xz -C /path/to/output

The validation and test archives can be extracted individually. Extraction is optional and requires additional disk space; the files in this repository are kept compressed; reconstruction reproduces the original bytes exactly.