HugMinjun/Blender6D-Track4D-v3

Blender6D Track4D v3 conditions — BF16 storage

Preprocessed conditioning data for 5,970 synthetic scenes and 23,880 directed camera pairs. Each scene supplies cam_00 → cam_01, cam_01 → cam_02, cam_02 → cam_03, and cam_03 → cam_00 at the same times.

BF16 revision, 2026-10-24

Memory and control latents are now stored as BF16 with lossless Zstandard compression, using bf16_dynamic_zstd_v1. Camera and geometry dtypes are preserved.

These files were converted from the prior INT8 release. The conversion stores exactly the BF16 values returned by the previous loader, verified bit-for-bit for every tensor. It does not recover precision lost during the original INT8 quantization and is not a fresh full-precision inference run. Per-sample precision_provenance records origin=dequantized_int8, the prior quantization errors, and original_precision_recovered=false. Source RGB, geometry, cameras, poses, captions, pair identities, and frame selection are unchanged.

The previous INT8 publication remains available at revision f3f92e5cf02a19b48a89bb58a7b4ee91bfc36ffe.

Contents

• 33 target RGB frame indices at 30 fps; input resolution 704 × 1280.
• Five latent anchors; BF16 memory shape [128, 5, 22, 40] and control shape [2, 128, 5, 22, 40], with SANA normalization.
• English captions, camera information, target indices, condition-image references, quality reports, object poses, and per-pair provenance.
• Metric geometry in metres: foreground depth uses a 0.001 conversion from source millimetres; background geometry comes from the source Blender scene.
• All detected dynamic regions are excluded before static-memory construction.
• Cross-camera identity conditions; no bank-shuffle variants for Blender6D.

Raw RGB videos, original .blend files, external target latents, and text embeddings are not included. Those inputs must be supplied separately when required by training. SpatialVID and Stereo4D outputs are separate datasets.

Files

shards/train-00000.tar through shards/train-00046.tar contain checksum-verified per-pair conditions.zip, pairs-0000.jsonl, receipt.json, and bf16-storage.json under data/scene_XXXXXX__cam_XX-cam_XX/.

The top-level pairs.jsonl is portable: archive paths are relative to the download root and RGB references use a source-videos/ placeholder. Per-pair manifests and archive metadata preserve original producer paths as provenance. Use the top-level manifest after resolving paths for your server.

checksums.json records SHA-256 and byte counts for the publication files. provenance/ contains the preprocessing configuration, completion audit, and BF16 migration summary. Training checkpoints, execution logs, credentials, and rollback copies are excluded.

Download and prepare

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="HugMinjun/Blender6D-Track4D-v3",
    repo_type="dataset",
    local_dir="blender6d-track4d-v3-bf16",
)

Use Python 3.12+, or a Python 3.11 release supporting tar extraction filters:

python blender6d-track4d-v3-bf16/prepare_manifest.py \
  --root blender6d-track4d-v3-bf16 \
  --video-root /path/to/train_geometry_rgb_instance_motion_gmos_v2 \
  --verify-shards

The RGB root must contain scene_XXXXXX/cameras/cam_XX/rgb.mkv. The preparation script extracts the archives, resolves paths, and writes pairs.local.jsonl. It reports missing RGB videos without providing those videos.

Use a Track4D loader that recognizes bf16_dynamic_zstd_v1. In the prior training_preprocessing/compact.py, add this identifier to READABLE_FORMATS; the existing decoder already leaves BF16 tensors without _qscale_ unchanged. The included bf16_loader.py can also inspect individual samples independently:

import json
from bf16_loader import load_conditions

with open("pairs.local.jsonl") as f:
    row = json.loads(next(f))
sample = load_conditions(row["archive"], row["sample"])
assert str(sample["memory_latent"].dtype) == "torch.bfloat16"

This minimal reader requires torch, safetensors, and zstandard; it does not load RGB or create training targets. The initial condition latent must come from the external target's exact first latent according to external_target_contract.

Static memory correction (2026-10-25)

Use the static memory correction folder with both the original cross-camera release and temporal-same-camera-33/. It repairs 6,187 affected condition archives while preserving all original downloads. The corrected static hole-fill pyramid can materially change affected memory latents. Other condition tensors, targets, captions and RGB references are unchanged.

Download holefill-fix-20261025/ and run its apply_patch_manifest.py on each base release's prepared local manifest. Use the resulting corrected manifest for training; do not concatenate old and corrected manifests. The correction contains only replacement archives, so existing base downloads can be reused.