Motion Synthesis on Various Body Shapes

We showcase results of our model ShapeMove when it is applied to synthesize the same actions for different body shapes. In the following videos, we use the same motion description with different shape descriptions to demonstrate how our method captures the natural variations in performing the same action by the different body types. The first row of video illustrates four sample body types. The Action buttons allow choosing and displaying different motions executed by these four body types.

Actions

Method ↑ back to top

In the inference phase (b), our model predicts motion tokens and the shape parameter from text inputs. We de-quantize these tokens using FSQ, and project into shape parameters with Projector. We concatenate Shape and Motion feature, and decode into the generated motion sequence with the Motion Decoder. Our model effectively synthesizes shape parameters and shape-aware motions reflecting the physical form and actions described in the input text.

Comparisons with Baselines for Given Body Shapes ↑ back to top

We showcase comparative results here among all of the baselines, including T2M-GPT [1], MotionDiffuse [2], MotionGPT [3], MLD [4], and MDM [5]. We show the ground truth motion and body shape in the first column for reference. This page contains 30 samples categorized by different actions. Selecting a particular Action, e.g., 'Run', shows a list of sample videos corresponding to the selected action type. Each individual video in this list can then be selected and viewed.

Actions

Videos