HugMinjun/Syn4D-Track4D-v3-training

Syn4D Track4DGEN v3 training caches

The GPU stage's outputs for HugMinjun/Syn4D-Track4D-v3: memory / control conditions, target latents and caption embeddings for SanaTrack training, made from that dataset's 5,817 windows of Syn4D V1 (revision bb08d13f).

This is a derivative of Syn4D, not an official Syn4D release. Please cite Syn4D (arXiv:2605.05207, project page) when you use it.

Contents

Pairs	12,051 memory → target pairs of 33 frames at 15 fps (manifests/pairs.jsonl)
Clips	5,682 windows
Conditions	61,374 files: per clip, pairs/conditions.zip with each pair's memory latent, control latents, cameras and object paths
Target latents	12,051 files: SANA-WM LTX-2 VAE latents [128, 5, 22, 40] of the 33-frame target chunk at 704 × 1280; the first temporal slice is the condition-image latent
Text	5,683 files: Gemma-2-2B-it caption embeddings [1, 300, 2304] through the SANA-WM text pipeline, plus negative.safetensors
Storage	bf16 latents and embeddings (no int8 quantization)
Size	38.4 GiB in uncompressed tar shards

The condition images are not duplicated here. They are decoded from the windows' clip.mp4, which you get from HugMinjun/Syn4D-Track4D-v3.

Layout

manifests/pairs.jsonl          the training manifest (one row per pair)
manifests/loader.json          SanaTrack loader entry (paths of the producing server, see below)
manifests/plan.json            the export plan
manifests/selection_report.json, index_b.jsonl   instance selection report and the GPU stage's index
shards/conditions-<h>.tar      v3/conditions/clips/<h>*/...   (h = first hex digit of the clip directory)
shards/targets-<h>.tar         v3/targets/<h>*/...
shards/text.tar                v3/text/...
SHA256SUMS

Tar members are relative to the Syn4D root. Extract every shard into the same root as the CPU release:

ROOT=/data/Syn4D
# 1. clips: the CPU release (tools/syn4d/hf_download.py in Track4DGEN), giving $ROOT/syn4d_tracks/*/*/clip.mp4
# 2. this release
hf download HugMinjun/Syn4D-Track4D-v3-training --repo-type dataset --local-dir "$ROOT/training-release"
(cd "$ROOT/training-release" && sha256sum -c SHA256SUMS)
for t in "$ROOT"/training-release/shards/*.tar; do tar -xf "$t" -C "$ROOT"; done
cp "$ROOT"/training-release/manifests/pairs.jsonl "$ROOT"/v3/conditions/pairs.jsonl
# 3. the paths in pairs.jsonl / loader.json are the producing server's (/workspace/dataset/Syn4D): map them
python scripts/prepare_datasets.py --out configs/main --dataset "$ROOT"/training-release/manifests/loader.json \
  --map /workspace/dataset/Syn4D="$ROOT"

How it was made

Track4DGEN tools/syn4d/run_gpu.sh (branch datasets/syn4d, commit b905d1c), steps 2–5 of tools/syn4d/GPU_GUIDE.md:

1. Export: memory / control conditions of every pair kept by the instance selection (dynamic_first_v3, boxes gravity_median_box_v1 on Syn4D's ground-truth gravity), scripts/run_v3_parallel.py --windows --storage-format bf16.
2. Target latents: tools/sanatrack/build_pair_targets.py (SANA-WM bidirectional VAE, e96271d).
3. Text: tools/sanatrack/build_text_cache.py (gemma-2-2b-it, 569d9809).
4. Completeness: every pair of pairs.jsonl has its target latent and its clip's text embedding.

Produced on 4 × TITAN RTX (Turing, no native bf16). Two changes to b905d1c were used:

• The VAE and Gemma computed in fp32 (T4D_COMPUTE_DTYPE=fp32 in ltx2_vae.py and tools/sanatrack/sanatrack/sana.py); outputs are cast to bf16 for storage as usual. Values can differ slightly from a bf16-compute run.
• Gravity vectors are passed to the tracking worker in positional notation (scripts/v3_track_worker.py): values such as -8.2e-05 or -0.0 were otherwise parsed as command-line options and stopped the export. The values are unchanged.

License and third-party assets

Released under CC BY 4.0, like Syn4D. Syn4D's LICENSE, LICENSES/ and THIRD_PARTY_NOTICES.md are included unchanged and apply as they do to Syn4D. Every sequence depends on BEDLAM2 bodies: no BEDLAM2 data is redistributed here, and you must obtain BEDLAM2 from its official source and agree to its license before using these sequences. No rights to third-party assets (Fab / Unreal scenes, Objaverse objects, Google Scanned Objects) are granted.