Magpie
Blender
Engine
UnrealZoo
YouTube
CARLA
WildWorld
OmniWorld
136K
clips
243 h
of video
21.5M
frames
84.1M
caption words
Collect
8 sources, 6 conditions
Align
158 frames at 24 fps
Check
sync and edge recall
Caption
scene + definitions
46.2K
YouTube
30.6K
Magpie
25.3K
CARLA
13.2K
Engine
8.5K
WildWorld
6.7K
Blender
4.1K
OmniWorld
1.8K
UnrealZoo
structure + first
structure
first frame
text
Time of day
day 66%
dusk 17% · night 12% · dawn 4%
Scene (share of clips)
indoor
15%
mountain
21%
waterside
21%
road
39%
city street
49%
Scale and Construction
Inputs per Source
Content Diversity
Caption Vocabulary
Materials
metal
wooden
stone
leather
concrete
Light, weather
shadows
haze
sunlight
glowing
fog
rain
Objects
trees
buildings
windows
water