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.

Camera trajectories come from MegaSaM (DepthAnything → UniDepth → DROID-SLAM, whole-person masked) and are metric — in metres, so translations are directly comparable across clips without per-clip rescaling.

Contents

config	clips	video hours	size on disk
parts_001-030	8,989	~12 h	~10.5 GB

parts_031-040 will be added as a separate config when its MegaSaM reconstruction finishes; existing configs will not change.

Every clip in the parquet has a matching record in the annotations NPZ and vice versa — the two are the same 8,989 clips, in the same order.

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_030)
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) w2c matrices, metric (metres)
camera_extrinsics_T	int32		number of keyframes T (72–130, median 106)
camera_extrinsics_kf_inds	int32	(T,)	source-frame index of each keyframe
face_bbox, lip_bbox	float32	(4,)	first-frame face / lip bbox [x0, y0, x1, y1] in pixels
face_conf, lip_conf	float32		first-frame detection confidence
face_bbox_seq_flat	float32	(T·4,)	flattened (T, 4) per-frame speaker face boxes
lip_bbox_seq_flat	float32	(T·4,)	flattened (T, 4) per-frame speaker lip boxes
conf_seq	float32	(T,)	per-frame detector confidence
match_seq	int8	(T,)	1 = frame matched the TalkNet speaker, 0 = no match (boxes blank)
speak_conf_seq	float32	(T,)	TalkNet per-frame speaking probability of the boxed person
track_seq	int16	(T,)	which TalkNet track the box follows at this frame (-1 = blank)
n_visible_tracks	int16		number of visible face tracks that competed for "speaker"
speaker_switches	int16		how many times the boxed identity changes mid-clip (0 = single face throughout)
align_factor, camera_scale	float32		align_factor = 1.0 — poses are already metric
vtss_score	float32		camera-motion filter score
long_caption, short_caption	string		Gemini-generated scene descriptions
dataset_source	string		always "OpenHumanVid"

The *_flat columns are flattened so parquet can store them as flat lists — reshape on read:

import numpy as np
T   = row["camera_extrinsics_T"]
ext = np.asarray(row["camera_extrinsics_flat"], dtype=np.float64).reshape(T, 4, 4)   # w2c, metres
lip = np.asarray(row["lip_bbox_seq_flat"],      dtype=np.float32).reshape(T, 4)      # pixels

face_bbox_seq_flat[k] pairs with camera_extrinsics_flat[k] and with source frame camera_extrinsics_kf_inds[k]. Frames where match_seq[k] == 0 have blank (zero) boxes — 90.9 % of the 935,004 keyframes (849,813) carry a TalkNet-speaker match.

Multi-speaker clips

The speaker is chosen per frame: at each keyframe the box follows whichever visible face has the highest TalkNet speaking probability (with hysteresis so it does not flicker at turn boundaries). In a two-person conversation the box therefore follows the turn — which is what you want for lip-audio alignment, but it means the boxed identity can change mid-clip.

If your model expects one face per sample, filter on speaker_switches == 0 (807 of 8,989 clips, 9 %, have at least one switch). n_visible_tracks == 1 (the majority of clips) can never switch.

Pipeline summary

1. Source: OpenHumanVid parts 001-030 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. MegaSaM metric SLAM — DepthAnything → UniDepth → DROID-SLAM with whole-person masks, so the tracker ignores human motion; output is metric c2w, stored here inverted to w2c.
8. Camera-motion + trajectory-roughness filter — keep clips that actually move (translation and rotation above the corpus median) and whose trajectories are not in the worst decile of any roughness metric (detour, curvature, absolute/relative lurch).
9. Per-frame speaker face+lip tracks — TalkNet best-speaker track joined to MediaPipe FaceMesh by timestamp and IoU.

Loading

from datasets import load_dataset
import numpy as np

ds = load_dataset("Haosonnn/OpenHumanVid-Talking", "parts_001-030", 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)  # w2c, metres
    # ...

Annotations NPZ (for the training pipeline)

annotations/train_data_openhumanvid_megasam_001-030_w2c.npz packages the same 8,989 clips as the parquet, without the video bytes — np.load(..., allow_pickle=True)["arr_0"] is a list of 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 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)	w2c keyframe matrices, metric (metres), already reshaped
camera_extrinsics_kf_inds	int32 (T,)	source-frame index of each keyframe
face_bbox, lip_bbox	float32 (4,)	first-frame bbox in pixels
face_conf, lip_conf	float32	first-frame detection confidence
face_bbox_seq, lip_bbox_seq	float32 (T, 4)	per-frame speaker boxes (already shaped, unlike the parquet's flat columns)
conf_seq	float32 (T,)	per-frame detector confidence
match_seq	int8 (T,)	1 = TalkNet-speaker match, 0 = none
speak_conf_seq	float32 (T,)	TalkNet per-frame speaking probability of the boxed person
track_seq	int16 (T,)	TalkNet track the box follows at this frame (-1 = blank)
n_visible_tracks, speaker_switches	int16	competing face tracks; mid-clip identity changes
align_factor, camera_scale, vtss_score	float32	align_factor = 1.0 (already metric)
video_width, video_height	int32	resolution

Join the NPZ to the parquet on key = basename of video_path without .mp4.

camera_extrinsics[0] ≈ I — the first camera is anchored at the origin, up to bundle-adjustment refinement (median deviation 5e-4).

Conventions

• Camera convention: OpenCV axes — R[:,0]=right, R[:,1]=down, R[:,2]=forward.
• Extrinsics are stored world-to-camera (w2c) in both the parquet camera_extrinsics_flat and the NPZ camera_extrinsics. Invert for camera-to-world: c2w = np.linalg.inv(w2c). (Corrected 2026-08-06 — READMEs before that date described these as c2w. The stored data was always w2c; only the documentation was wrong. Check any code written against the old wording.)
• Trajectories are metric, in metres, with align_factor = 1.0. No per-clip rescaling needed.
• 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 (MonST3R poses).

• 2026-07-18: added annotations/train_data_openhumanvid_monst3r_001-040.npz.

• 2026-08-06: annotations switched to MegaSaM (8,989 clips, parts 001–030, metric, per-frame speaker face/lip); MonST3R NPZ removed. Corrected the documented pose convention c2w → w2c (data unchanged).

• 2026-08-08: video shards rebuilt against the MegaSaM set. The parts_001-040 config (32,176 clips, MonST3R poses, 34 GB) was removed and replaced by parts_001-030 (8,989 clips, MegaSaM metric poses, ~10.5 GB). Parquet and NPZ now describe an identical clip set, and the parquet gained the per-frame speaker tracks. 25,465 MonST3R-only clips are no longer distributed here; their blobs remain in this repo's git history. Also corrected the stated video hours — the old "86 h" was about 2× too high (the 32,176-clip set measured ~41 h; mean clip length is ~4.6 s, not ~9.6 s).

• 2026-08-09: active-speaker fix — per-frame face/lip boxes corrected in both the parquet shards and the annotation NPZ. Trajectories, captions, video bytes, clip set and clip order are unchanged.

  Three defects, all in how the speaker was chosen:

  1. Wrong speaker during turn-taking. The speaker used to be picked once per clip as the track with the highest mean sync_conf. In a two-person conversation the person who talks longest wins that average, so during the other person's turn the box sat on a silent face. Measured at 0.65 % of boxed frames, with 0.5 % of clips more than 30 % affected. Now chosen per frame.
  2. match_seq == 2. Parts 001-010 were built by an older extractor with a "top-score fallback" that boxed the highest-confidence face when the speaker failed to match, labelling it 2. Parts 011-030 had no such frames, so one file carried two different definitions. All 15,651 such frames are gone; match_seq is now strictly {0, 1} everywhere. Of those, 3,025 were a genuinely wrong person (multi-face frames); the other 12,626 were single-face frames and are now simply re-derived under the strict rule.
  3. Scalar face_bbox / lip_bbox provenance. In 345 records of parts 001-010 these held values from a separate first-frame extraction rather than face_bbox_seq[0]. They are now true frame-0 aliases in all 8,989 records.

  New fields: speak_conf_seq, track_seq, n_visible_tracks, speaker_switches. If your model needs one identity per sample, filter speaker_switches == 0 — see Multi-speaker clips above. Speaker-matched keyframes: 850,238 → 849,813 of 935,004 (90.9 %), now on the correct face.