OpenHumanVid-Talking — camera-controlled, audio-conditioned talking-head videos

This dataset is a filtered, re-packaged subset of OpenHumanVid, curated for training camera-controlled, audio-conditioned talking-head video generation models. Each clip is a short talking-head segment where audio, camera trajectory, and portrait are aligned frame-by-frame.

Contents

config	clips	video hours	size on disk
parts_001-040	32,176	~86 h	~34 GB

More parts will be added as new configs (parts_041-050, …) without changing existing ones.

Per-clip fields

Each parquet row is one talking-head segment with:

field	dtype	shape	meaning
key	string		<parent_cid>_seg{NN} unique clip id
part	string		source part directory (part_001 … part_040)
video	binary		H.264 + AAC mp4 bytes (audio baked in)
video_width, video_height	int32		resolution
camera_intrinsics	float32	(4,)	[fx, fy, cx, cy] in pixels
camera_extrinsics_flat	float64	(T·16,)	flattened (T, 4, 4) c2w matrices at keyframes
camera_extrinsics_T	int32		number of keyframes T
camera_extrinsics_kf_inds	int32	(T,)	source-frame index of each keyframe
face_bbox	float32	(4,)	first-frame face bbox [x0, y0, x1, y1]
lip_bbox	float32	(4,)	first-frame lip bbox [x0, y0, x1, y1]
face_conf, lip_conf	float32		first-frame detection confidence
align_factor, camera_scale	float32		trajectory scale calibration (monocular SLAM output; not metric)
vtss_score	float32		camera-motion filter score
long_caption, short_caption	string		Gemini-generated scene descriptions
dataset_source	string		always "OpenHumanVid"

To rebuild the extrinsics matrix:

import numpy as np
ext = np.asarray(row["camera_extrinsics_flat"], dtype=np.float64).reshape(row["camera_extrinsics_T"], 4, 4)

Pipeline summary

1. Source: OpenHumanVid parts 001-040 raw videos (~50k clips/part).
2. Motion top 10 % by upstream global_motion score.
3. VAD (Silero) — keep clips with speech; ≥ 10 % voiced.
4. Prefilter — FastSAM × MediaPipe-Selfie joint deciles to drop static / non-talking-head.
5. TalkNet ASD + MediaPipe-v3 — per-frame speaker bbox and face landmarks.
6. Speech-segment carve — TalkNet best-track speech islands padded ±0.5 s, ≥ 3 s.
7. MonST3R monocular SLAM — camera trajectory per keyframe (camera_extrinsics).
8. Camera-motion filter — keep clips where either
   • camera-center arc length is above the 50th-percentile of the corpus, OR
   • forward-direction cumulative angle is above the 70th-percentile.
9. First-frame face+lip bbox via MediaPipe.

Loading

from datasets import load_dataset

ds = load_dataset("Haosonnn/OpenHumanVid-Talking", "parts_001-040", split="train", streaming=True)
for row in ds:
    mp4_bytes = row["video"]                # H.264 + AAC
    intr      = row["camera_intrinsics"]    # [fx, fy, cx, cy]
    T         = row["camera_extrinsics_T"]
    ext       = np.asarray(row["camera_extrinsics_flat"]).reshape(T, 4, 4)  # c2w
    # ...

Annotations NPZ (for the training pipeline)

annotations/train_data_openhumanvid_monst3r_001-040.npz packages the same per-clip annotations as the parquet, but as a single NumPy object array — np.load(..., allow_pickle=True)["arr_0"] is a list of 32,176 per-clip dicts. This is the form consumed directly by the DiffSynth-InContext-Control training code (precompute_openhumanvid.py / LoadOpenHumanVidCond); use it if you want the poses / bboxes / captions without decoding the parquet video shards.

key	dtype / shape	meaning
dataset_source	str	always "OpenHumanVid"
video_path, audio_path	str	clip path relative to an OpenHumanVid root (audio is baked into the same mp4)
short_caption, long_caption	str	Gemini scene descriptions
camera_intrinsics	float32 (4,)	[fx, fy, cx, cy]
camera_extrinsics	float64 (T, 4, 4)	c2w keyframe matrices, already reshaped (T ≈ 20, variable)
camera_extrinsics_kf_inds	int32 (T,)	source-frame index of each keyframe (interpolate to your sampled frames)
face_bbox, lip_bbox	float32 (4,)	first-frame bbox [x0, y0, x1, y1] in pixels (normalize by video_width/video_height)
face_conf, lip_conf	float32	first-frame detection confidence
align_factor, camera_scale, vtss_score	float32	scale calibration / motion score (align_factor is unused for these monocular MonST3R poses)
video_width, video_height	int32	resolution

import numpy as np
ann  = np.load("annotations/train_data_openhumanvid_monst3r_001-040.npz", allow_pickle=True)["arr_0"]
clip = ann[0]                       # dict of per-clip annotations
ext  = clip["camera_extrinsics"]    # (T, 4, 4) c2w keyframes; kf_inds gives their frame indices

The derived npz/metadata are also mirrored (with the RealCam-Vid camera subset) in Haosonnn/wan22-incontext-control-data.

Conventions

• Camera convention: OpenCV c2w. R[:,0]=right, R[:,1]=down, R[:,2]=forward.
• Trajectory scale is per-clip and not metric (monocular SLAM). Use align_factor / camera_scale if you need to normalize across clips.
• Audio is embedded in the mp4 (AAC track); no separate audio files.

License and citation

Distributed under CC-BY-NC-4.0 — non-commercial use only. This is a derivative of OpenHumanVid; please respect the source dataset's terms and cite the original work.

@article{openhumanvid,
  title  = {OpenHumanVid: A Large-Scale High-Quality Dataset for Enhancing Human-Centric Video Generation},
  author = {DeepGlint},
  year   = {2024}
}

Changelog

• 2026-07-15: initial release, parts_001-040, 32,176 clips.
• 2026-07-18: added annotations/train_data_openhumanvid_monst3r_001-040.npz (MonST3R poses + face/lip bbox + captions packaged for the training pipeline).