Select avatar parts and moments with natural language.
HumanSkill-4D uses a pretrained multimodal large language model (MLLM) to select body parts, clothing, and motion intervals without additional training.
Method
How It Works
Gallery · Where
3D Referring Segmentation
ZJU-MoCap 377
Drag to rotate · Scroll to zoom Swipe sideways to rotate
Example queries
Choose a query to view the selection.
Try the same query on another avatar
Gallery · Where × when
4D Referring Segmentation
Where Body part or garmentWhen Motion condition
Example queries
Choose a query to view the selection.
Try the same query on another person
Benchmark