HugMinjun/PointOdyssey-Track4D-v3-training

PointOdyssey Track4DGEN v3 training caches

The GPU stage's outputs for HugMinjun/PointOdyssey-Track4D-v3 (the CPU release cpu-gt-v1): memory / control conditions, target latents and caption embeddings for SanaTrack training, made from PointOdyssey (aharley/pointodyssey, revision 4a2ec949, train split).

This is a derivative of PointOdyssey, not an official release. Upstream dataset terms apply; no new license is granted here. Please cite PointOdyssey when you use it.

Contents

Pairs	1,912 memory → target pairs of 33 frames at 15 fps (manifests/pairs.jsonl)
Clips	105 sequences
Conditions	9,560 files: per pair directory, pairs/conditions.zip with each pair's memory latent, control latents, cameras and object paths
Target latents	1,912 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	106 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	5.1 GiB in uncompressed tar shards

The condition images are not duplicated here. They are decoded from the lossless RGB cpu-gt-v1/rgb/*.mkv, which you get from HugMinjun/PointOdyssey-Track4D-v3.

Layout

manifests/pairs.jsonl            the training manifest (one row per pair)
manifests/plan.json              the export plan
manifests/targets-selected.jsonl the targets the caches were built from
shards/clips-<h>.tar             v3/clips/<h>*/...     (h = first hex digit of the pair directory)
shards/targets-<h>.tar           v3/targets/<h>*/...
shards/text.tar                  v3/text/...
SHA256SUMS

Tar members are relative to the PointOdyssey root. Put the CPU release in $ROOT/release and extract every shard into $ROOT:

ROOT=/data/PointOdyssey
# 1. lossless RGB: the CPU release, giving $ROOT/release/cpu-gt-v1/rgb/*.mkv
hf download HugMinjun/PointOdyssey-Track4D-v3 --repo-type dataset --local-dir "$ROOT/release"
# 2. this release
hf download HugMinjun/PointOdyssey-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 "$ROOT"; done
# 3. pairs.jsonl holds the producing server's absolute paths (/workspace/dataset/PointOdyssey): map them
sed "s#/workspace/dataset/PointOdyssey#$ROOT#g" "$ROOT"/training-release/manifests/pairs.jsonl > "$ROOT"/v3/pairs.jsonl

Training reads $ROOT/v3/pairs.jsonl, with target latents in $ROOT/v3/targets and text in $ROOT/v3/text.

How it was made

Track4DGEN tools/pointodyssey/README.md "Receiver: exact continuation commands" (commit 3044edaf):

1. handoff.py localize of the CPU release.
2. Export: memory / control conditions of every selected pair (dynamic_first_v3, boxes gravity_median_box_v1), scripts/run_v3_parallel.py --windows --storage-format bf16. All 681 windows, 0 failures.
3. Target latents: tools/sanatrack/build_pair_targets.py (SANA-WM bidirectional VAE, e96271d).
4. Text: tools/sanatrack/build_text_cache.py (gemma-2-2b-it).
5. Completeness: every pair of pairs.jsonl has its target latent and its clip's text embedding, and the negative embedding exists.

Produced on 4 × TITAN RTX (Turing, no native bf16). Two changes to 3044edaf 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), so values such as -8.2e-05 or -0.0 are not parsed as command-line options. The values are unchanged.