MiniMax-H3 Human–Scene Interaction Dataset

894 paired samples of video + text prompt + 3D human motion (SMPL) for training joint video–motion generation models. Each clip shows a single person performing a language-specified human–scene interaction (walking to a bench and sitting down, leaning against a stone, etc.) under a locked-off static camera, with the full body visible in frame at all times.

Videos were generated with MiniMax-H3 (first-frame image + prompt), and per-frame SMPL parameters were extracted with CameraHMR.

Contents

video/         894 clips, 1024x576 @ 24 fps, ~120-150 frames each (mp4/h264)
hmr/           894 per-clip SMPL sequences extracted by CameraHMR (npz)
prompt.json    list of {"name": "<clip>.mp4", "prompt": "<text>"} for all 894 clips

hmr/<name>_hmr.npz fields

key	shape	description
global_orient	(T, 3)	SMPL global orientation, axis-angle, camera frame
body_pose	(T, 23, 3)	SMPL body pose, axis-angle
betas	(T, 10)	SMPL shape parameters (per frame)
cam_t	(T, 3)	root translation in camera coordinates (meters)
valid	(T,)	per-frame detection validity mask
cam_int	(3, 3)	camera intrinsics K, defined on the 1024x576 image plane
fps	()	frame rate (24.0)

T equals the number of video frames. The camera is static, so cam_int and the world-to-camera transform are constant within a clip; motion is metrically grounded in the camera frame of the clip.

Naming convention

<id>_<scene>_<object/variant>_<subject>_<emotion>.mp4, e.g. 0000_platform_long_mm_happy.mp4. Subject codes combine gender/age (mm/mw/ym/yw/om/ow/tg ...); the trailing tag is the emotion condition (happy / low).

Loading example

import json
import numpy as np

prompts = {e["name"]: e["prompt"] for e in json.load(open("prompt.json"))}
d = np.load("hmr/0000_platform_long_mm_happy_hmr.npz")
print(prompts["0000_platform_long_mm_happy.mp4"][:80])
print(d["body_pose"].shape, d["cam_t"].shape, d["cam_int"])