NL3D/NL3D-Synth-Train

NL3D Synthetic Training Data

This repository contains synthetic NL3D training artifacts. The notes below document the cold backup stored under dir1/backup_shards_v1/; other packaged or composed dataset artifacts in this repository may use a different layout.

Composed training backups

The composed training set is stored in eight independently packaged batches. Together they contain 5,335 scenes in 42 uncompressed TAR shards, totaling 2,074,187,438,080 bytes.

Batch	Repository location	TAR shards	Scenes	TAR bytes
compose1	archives/compose1_shards50g_20260517/	5	648	251,517,511,680
compose2	archives/compose2_shards50g_20260515/	4	499	201,590,394,880
compose3	archives/compose3_shards50g_20260516/	4	500	193,708,922,880
compose4	archives/compose4_shards50g_20260517/	5	688	267,236,935,680
compose5	archives/compose5_shards50g_20260518/	4	500	193,601,300,480
compose6	archives/compose6_shards50g_20260518/	4	500	193,650,227,200
compose7	shards/compose7-*.tar	8	999	386,219,550,720
compose8	shards/compose8-*.tar	8	1,001	386,662,594,560

Each compose1-6 archive directory contains its own bundle.env, samples.tsv, shards.tsv, SHA256SUMS, and shards/ directory.

Compose7 and compose8 predate that archive layout and share the repository-level shards/ directory. The unprefixed root control files (bundle.env, samples.tsv, shards.tsv, and SHA256SUMS) belong to compose8. Compose7 uses the separate root files compose7_bundle.env, compose7_samples.tsv, compose7_shards.tsv, and compose7_SHA256SUMS. Keeping these names distinct prevents one batch's manifests from replacing the other's.

To download and verify compose7:

hf download NL3D/NL3D-Synth-Train \
  --repo-type dataset \
  --include 'compose7_*' 'shards/compose7-*.tar' \
  --local-dir /data/jiachen/compose7_restore
cd /data/jiachen/compose7_restore
sha256sum -c compose7_SHA256SUMS

For compose8, download shards/compose8-*.tar together with the four unprefixed root control files, then run sha256sum -c SHA256SUMS from the download root.

dir1/ cold backup

The canonical files selected from the local output_1, output_2, and output_3 directories are stored as scene-aligned, uncompressed TAR shards. The original individual-file tree has been removed from the current main revision after all archive files were verified remotely.

Item	Count or size
TAR shards	26
Canonical source files	842,008
Scene directories	3,315
Complete scenes	3,314
Approximate source size	1.285 TB
Total TAR size	1,285,341,429,760 bytes

Each scene is wholly contained in one shard. TAR members start with dir1/, so extracting all shards into one destination reconstructs the unified canonical scene tree. The original output_1/output_2/output_3 membership of every scene is retained in samples.tsv.

The backup includes canonical PNG render passes, scene metadata, and geometry files. It excludes simulation caches, logs, EXR render intermediates, and other non-canonical working artifacts.

Repository layout

dir1/backup_shards_v1/
├── README.md
├── bundle.env
├── samples.tsv
├── shards.tsv
├── SHA256SUMS
└── shards/
    ├── output_1-00000.tar ... output_1-00008.tar
    ├── output_2-00000.tar ... output_2-00008.tar
    └── output_3-00000.tar ... output_3-00007.tar

A complete extracted scene normally has 254 files:

dir1/<group>/<scene_id>/
├── rgb/000.png ... 049.png
├── diffuse/000.png ... 049.png
├── glossy/000.png ... 049.png
├── transmission/000.png ... 049.png
├── mask/000.png ... 049.png
├── material.json
├── transforms.json
├── mesh_raw.ply
└── points_gt.ply

Known incomplete scene

The following source scene contains 252 canonical files and is missing mesh_raw.ply and points_gt.ply:

dir1/swamp/88903d19ef344680b64ced73b86cffc0

This is a known source-data omission, not archive corruption.

Restoring the backup

Downloading the repository with Git or Git LFS is not recommended. Download the backup directory with the Hugging Face CLI:

hf download NL3D/NL3D-Synth-Train \
  --repo-type dataset \
  --include 'dir1/backup_shards_v1/**' \
  --local-dir /data/jiachen/nl3d_restore_download

Verify all downloaded TAR files:

cd /data/jiachen/nl3d_restore_download/dir1/backup_shards_v1
sha256sum -c SHA256SUMS

Extract the unified dir1/ tree:

mkdir -p /data/jiachen/nl3d_restore
for shard in shards/*.tar; do
  tar -xf "$shard" -C /data/jiachen/nl3d_restore
done

For reproducible recovery, add --revision <commit_sha> to the download command. A complete download and extraction requires space for both the archives and extracted files at the same time; provision at least 2.6 TB, with additional headroom recommended.

Backup format note

The TAR layout is intended for reliable cold storage and restoration. It reduces the current repository tree from hundreds of thousands of individual paths to a small set of independently verifiable archives. It is not a WebDataset training layout.