---
license: cc-by-4.0
task_categories:
- image-to-video
language:
- en
tags:
- human-animation
- pose-guided
- dwpose
- reference-to-video
- wan2.2
- one-to-all-animation
pretty_name: HumanVid-Vertical OTA Stage1 Pack (poses + manifest, videos by URL)
size_categories:
- 1K<n<10K
---

# HumanVid-Vertical OTA Stage1 Pack

**5,828 single-person vertical (1080×1920) video clips with per-frame DWPose
annotations**, prepared as a Stage1 (reference/appearance) training pack for
[One-to-All-Animation](https://github.com/gaojunyao/One-to-All-Animation)-style
pose-guided human video generation on Wan2.2-TI2V-5B.

**Videos are NOT included.** Source videos come from Pexels via the
[HumanVid](https://github.com/zhenzhiwang/HumanVid) (CC-BY-4.0) URL release and
cannot be redistributed; this repo follows the same pattern as HumanVid itself:
we ship **URL lists + derived annotations + exact reproduction scripts**, and
you download the videos yourself (one command, ~1-2h, ~60GB raw).

## What's inside

| Path | Contents |
|---|---|
| `poses/poses_vertical_150f_*.tar` | 5,828 DWPose `.npy` files (one per clip, 150 frames each, ~2.9GB total) |
| `manifests/humanvid_vertical_stage1.csv` | Stage1 training manifest, **relative paths** (`videos/<name>.mp4`, `poses/<name>.npy`) |
| `urls/pexels-vertical-urls-new.txt` | 7,851 direct Pexels mp4 URLs (from HumanVid's updated lists) |
| `urls/pexels-horizontal-urls-new.txt` | 11,411 landscape URLs (not yet processed; included for completeness) |
| `scripts/` | The full pipeline: downloader, trimmer, DWPose extraction (CPU + GPU/onnx2torch), decord shim for aarch64 |

## Pose format

Each `.npy` is a length-150 object array; element `t` is the DWPose dict for
frame `t`:

```python
{
  "bodies": {"candidate": (18,2) float64 in [0,1] (x,y), "subset": (1,18), "score": (1,18)},
  "hands": (2,21,2), "hands_score": (2,21),
  "faces": (1,68,2), "faces_score": (1,68),
}
```

Coordinates are normalized to the frame; invisible keypoints are `-1`
(score < 0.3). This is exactly the object-array format the One-to-All
`bodydance_*` dataloaders consume via `pose_path`.

## Rebuild the videos (required before training)

```bash
# 1. Download the source clips from Pexels (they are NOT in this repo)
LIST=urls/pexels-vertical-urls-new.txt OUT_DIR=videos_raw \
  bash scripts/download_humanvid_pexels.sh

# 2. Trim every clip to its first 150 frames (MUST match pose length —
#    the One-to-All dataset samples the reference frame from the PHYSICAL
#    video length, so untrimmed videos overflow the pose arrays)
SRC_MANI=manifests/humanvid_vertical_stage1.csv OUT_DIR=videos \
  bash scripts/trim_videos.sh   # edit FF= to your ffmpeg path

# 3. Untar poses
mkdir -p poses && for t in poses/poses_vertical_150f_*.tar; do tar -xf "$t" -C poses; done

# 4. Point the manifest's relative paths at your absolute location
#    (or run training from this directory).
```

A small number of URLs may have gone stale on Pexels since collection
(28/7,851 were dead at collection time); drop missing rows from the manifest.

## How it was built

1. **Download**: HumanVid vertical URL list (7,851), 7,823 fetched.
2. **Filter + DWPose** (per clip, first 150 frames = 6s @25fps):
   single-person filter (YOLOX person detector, ≥80% single-person frames),
   scene-cut filter (≤1 cut), DWPose wholebody per frame.
   Kept 5,828 (74.5%); rejects were ~99% multi-person.
3. **Trim**: kept clips physically cut to 150 frames (`-frames:v 150`, h264 crf18, audio dropped).
4. **Dataloader audit**: the pack was verified through the real One-to-All
   Stage1 dataloader (`bodydance_refmask` + `Collate`): reference-image CFG
   dropout rate matches config, pose-on-video overlays visually aligned,
   OCR masks all-keep, face-mask coverage mean 4.2% (4/24 sampled clips have
   no detectable face and fall back to uniform loss weight).

Extraction ran on a DGX Spark (GB10, aarch64): DWPose ONNX models converted to
torch via onnx2torch (no aarch64 onnxruntime-gpu exists); GPU outputs validated
against CPU onnxruntime (mean keypoint diff 0.02px). See `scripts/gpu_dwpose.py`.

## Intended use & licensing

- **Annotations + manifest + scripts (this repo): CC-BY-4.0.**
- **Videos**: Pexels License (free to use, no redistribution as stock) — hence
  URL-only distribution. Review the [Pexels license](https://www.pexels.com/license/)
  before commercial use.
- URL lists derive from HumanVid (CC-BY-4.0). If you use this pack, please cite
  HumanVid (Wang et al., NeurIPS D&B 2024), DWPose (Yang et al., ICCV 2023),
  and One-to-All-Animation.

## Caveats

- Composition is Pexels stock footage: clean, well-lit, mostly moderate motion.
  It complements — but does not replace — dance/high-dynamics data (AIST++,
  Open-HyperMotionX) or face-heavy data (TalkVerse, VividHead) in a training mix.
- Captions are a fixed neutral sentence (`qwen_caption` column); the intended
  Stage1 recipe trains with `--cfg 1.0` (text always dropped), so captions are
  placeholders.
- 150 frames @ 25fps ≈ 6.0s per clip; `fps` column is 25, train-time resampling
  to 16fps is handled by the loader (`--train_fps 16`).