Cosmos-Xenna (Data Curation)

Xenna is NVIDIA's 7-stage video data curation pipeline — the same one used to prepare Cosmos training corpora.

Stages

1. split — shot detection & clipping
2. transcode — normalize resolution/fps/codec
3. crop — detect and remove borders/letterboxing
4. filter — drop low-quality / duplicate / unsafe clips
5. caption — auto-label with VLM captions
6. dedup — embedding-based near-dupe removal
7. shard — pack into WebDataset/TAR shards

Agent tool

cosmos_curate(
    input_dir="./raw_videos",
    output_dir="./outputs/curated",
    stages="all",                     # or "split,transcode,filter"
    num_workers=8,
)

CLI

just curate ./raw_videos ./outputs/curated all 8
just curate ./raw_videos ./outputs/curated "split,caption" 4

Typical pipelines

Full pipeline

just curate ./raw ./curated all 16

Captioning only (already cleaned)

just curate ./cleaned ./captioned caption 8

Test run (subset of stages)

just curate ./test_videos ./out "split,transcode" 4

Ray cluster

Xenna uses Ray for distributed work. The num_workers arg controls parallelism per node. For multi-node, configure Ray in your Cosmos-Xenna clone (see COSMOS_XENNA_REPO).

Output layout

outputs/curated/
├── shards/
│   ├── train-000000.tar
│   ├── train-000001.tar
│   └── ...
├── captions/
│   └── captions.parquet
└── stats.json

References

• Data curation overview
• Cosmos-Xenna repo