The SDG-SynHuman is a large-scale synthetic video dataset of digital humans rendered in diverse indoor and outdoor 3D environments. The dataset contains 236,937 clips, totaling approximately 5,841 hours of video, and is designed to support training and post-training of NVIDIA Cosmos world foundation models and related physical AI research.

Each sample is a temporally coherent 60-120 second video clip rendered at 1080p and 30 fps. Clips contain multiple digital humans performing animation sequences in a sampled 3D environment with a controlled camera trajectory. The dataset spans 4,050 digital human assets, 8,184 unique animations, 198 indoor environments, 200 outdoor city environments, and 14 camera-motion scenarios, providing broad variation in human appearance, motion, scene context, lighting, and camera behavior.

The dataset is intended as a controllable synthetic supplement to real-world human video data for applications such as world-model pretraining and post-training, camera-motion generalization, depth and geometry-aware learning, human-scene interaction modeling, and physical AI research.

This dataset is fully synthetic. It contains no real-world imagery, no personally identifiable information, and no real human subjects. All content was generated procedurally using NVIDIA's Synthetic Data Generation pipeline with synthetic digital humans and synthetic or simulation-ready 3D scene assets.

This dataset is ready for commercial/non-commercial use.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

Creation Date: 2026-04-28

Version

PhysicalAI-WorldModel-Synthetic-Digital-Human-Scenes v1.0.0

License/Terms of Use:

This dataset is released under the OpenMDW1.1.

Intended Usage:

SynHuman is intended for AI/ML researchers and developers working on world foundation models, video generative models, human-centric video understanding, camera-control modeling, and physical AI.

The dataset is designed to serve as a controllable synthetic supplement to real-world human video data, especially for tasks that benefit from deterministic camera and geometry supervision.

Typical use cases include:

pretraining and post-training video world models
training or evaluating camera-motion generalization
depth- and geometry-aware learning with paired RGB, metric depth, and camera calibration
human-scene interaction modeling
studying the role of synthetic data in physical AI systems

This dataset is not intended to replace real-world validation. Models trained with SDG-SynHuman should be evaluated on representative real-world data before use in production systems.

It should not be used for biometric identification, real-person recognition, surveillance targeting, or inferring sensitive attributes of real individuals.

Dataset Statistics

This dataset contains 236,937 samples, totaling 5,841 hours of video. Clips are rendered at 1080p and 30 fps, with durations between 60 and 120 seconds and an average duration of approximately 88.8 seconds.

This corresponds to roughly 631 million RGB frames, with paired metric depth frames and camera parameters generated at the same temporal resolution.

The dataset spans:

4,050 unique digital human assets
8,184 unique animations
198 indoor environments
200 outdoor city environments
14 camera-motion scenarios

Each scenario contains one 3D environment, one lighting condition, one camera trajectory, and one to nine digital humans, providing broad variation across human appearance, animation, scene context, and camera motion.

Primary Camera Motion

Distribution across 236,937 primary-motion occurrences.

Motion Type	# Scenes	Percentage	Description
static	71,006	29.97%	Camera remains fixed in position and orientation throughout the sequence
egocentric	33,070	13.96%	Camera behaves as the viewpoint of a character or agent
tracking	36,978	15.61%	Camera follows a subject while maintaining framing
flythrough	36,852	15.55%	Camera travels forward through the scene along a path
arc	37,166	15.69%	Camera moves in a curved arc around a target
zigzag	15,294	6.45%	Camera advances with alternating lateral motion
birdseye	6,571	2.77%	Top-down overhead camera perspective
Total	236,937	100.00%	---

Secondary Camera Motion

Secondary motions are compositional layers that may overlap temporally. Reported durations are therefore independent per-motion totals rather than shares of a fixed duration budget.

Motion Type	Active Time (hours)	Description
breathing	1,221.77	Gentle breathing-like motion to mimic human respiration
drift	1,211.96	Slow subtle positional motion over time
dutch_angle	1,205.72	Rotation around the forward axis producing a tilted horizon
shake	1,197.87	Rapid positional and rotational handheld jitter
sway	1,191.28	Pendulum-like oscillatory motion with horizon rocking
zoom	1,184.09	Forward/backward motion without lens-property changes
crab	373.72	Lateral translation while maintaining viewing direction
tilt	195.33	Up/down rotational motion around the horizontal axis

Dataset Characterization

** Data Collection Method

[Synthetic] - [SDG-SynHuman is a fully synthetic dataset generated through NVIDIA’s Synthetic Data Generation pipeline. Scenarios are procedurally sampled from structured world configurations that define the environment, lighting, digital humans, animations, camera model, and camera trajectory. Each clip is rendered in a 3D simulation environment using synthetic digital humans and synthetic or simulation-ready scene assets, including internal NVIDIA environments, indoor environments adapted from the SceneSmith example-scenes dataset, and outdoor city environments generated with Blender’s CityGenerator plugin. No real-world video footage, real human-subject data, or real-world audio is used.]

** Labeling Method

[Synthetic] - [All annotations are synthetic and generated deterministically from the underlying USD scene graph during simulation. The release includes RGB video, metric depth, per-frame camera intrinsics and extrinsics, world configuration metadata, and asset-level metadata describing the environment, digital humans, animation assignments, placements, and camera behavior. No human labeling is involved.]

Dataset Format

This dataset is delivered in 1,215 tar shards found in \path{shards/}. Each tar bundles 200 samples. Each sample is identified by its UUID.

Modality: Video (RGB and depth) with structured annotations.

Format: RGB video is encoded as MP4 (H.264) at 1080p / 30 fps. Depth video is encoded as MKV (FFV1, lossless) at 1080p / 30 fps.

Clip duration: Each clip is approximately 60-120 seconds long.

Packaging: Each sample is associated with a UUID and includes rendered RGB video, depth output, per-frame camera parameters, and scene metadata.

Per-sample files include:

video/uuid.mp4 - RGB video rendered at 1080p and encoded as H.264.
depth/uuid.mkv - Metric depth rendered at 1080p as distance to the image plane, encoded losslessly with FFV1.
depth/uuid_depth.json - Companion depth metadata, including depth range, resolution, frame rate, and data type for metric reconstruction.
meta/uuid_camera.json - Per-frame camera intrinsics and extrinsics.
meta/uuid.json - Scene configuration, including environment, lighting, digital-human agents, camera setup, and animation task definitions.
description/uuid.json - Asset metadata, including spawned asset inventory, digital humans, motions, placements, and provenance information.

Dataset Quantification

Record Count: 236,937 records, totaling approximately 5,841 hours of video.

Feature Count (per record): RGB and depth videos, depth companion metadata (depth range, resolution, frame rate, and dtype), per-frame camera intrinsics and extrinsics, world configuration metadata, and asset configuration metadata.

Total Data Storage: Approximately 155.98 TB.

Reference(s):

HuggingFace:
https://huggingface.co/datasets/nvidia/PhysicalAI-SDG-SynHuman

Upstream environment asset reference:
https://huggingface.co/datasets/nepfaff/scenesmith-example-scenes

Blender CityGenerator plugin (version 2.4):
https://superhivemarket.com/products/the-city-generator

Related announcement — NVIDIA Physical AI Dataset:
https://blogs.nvidia.com/blog/open-physical-ai-dataset/

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.