One Kick, Six Motion Models

One generated right-foot kick, six reconstruction candidates. Our visual review preferred SAM 3D Body, followed by GVHMR and PromptHMR. MotionBERT was weakest on this clip; WHAM and PhysPT were left unranked.

The source action

MiniMax H3 generated reference: 3 seconds · 72 frames · 24 fps · 1,024 × 576. A fixed camera captures preparation, left-leg support, a right-foot strike, follow-through, and recovery. This is an approved action candidate, not measured motion capture.

The visible strike window is frames 41–42. The frame labels describe the images, not a measured three-dimensional contact point.

Compare the motion yourself

Open the interactive comparison at full width.

Switch views, step through frames, or enlarge two models. Blue is the left side; orange is the right. Pelvis alignment changes the display only. A reconstructed side view is another view of an estimate, not another camera observation.

The quick replay

All six native estimates use the same source timeline. No extra foot locking or smoothing was added. PhysPT's native 20 fps output is mapped to 24 fps; the last frame holds its final native sample. MotionBERT is a joint-position baseline.

Depth, scale, shoe contact, and support-foot sliding remain unmeasured. Next: map SAM 3D Body onto Kickshatter's actual player_v36 rig in an isolated preview. This single-clip preference is not a general model ranking.

Source video · Comparison video

Models: GVHMR · PromptHMR · WHAM · SAM 3D Body · MotionBERT · PhysPT.