Motion Editing · Style and Content

Exact-source motion editing on held-out PerMo pairs. Style editing changes motion character while preserving action; Content editing changes action while preserving style. Every ranked method is measured with the same ten-metric Universal SMPL-H protocol.

Table 6.c · full protocolInspect all motions

63 Style cases67 Content casesExact source + instruction30 fps physical auditUniversal SMPL-H uTMR

Ranked generated methods

Held-out cases in active track

Shared reported metrics

M2M R@3

Default display rank

Aggregate Comparison

Metric

Measured values

Normalized profile

100 denotes the best ranked value on each non-reference axis.

Results

No custom score and no native-input column: all ten paper metrics remain visible in one table.

Click any metric header to sort.

Best ranked methodSecond ranked methodREF = source inputDIAG = different condition contract

Full Qualitative Audit

Source, paired target, every reproducible baseline and MotionCanvas are synchronized case-by-case in Three.js.

Held-out case

Camera

Loading case index…

0.00s0.00s

Loading SMPL-H assets…

Every panel uses the same SMPL template, skinning weights, camera, lighting, and synchronized timeline. HML263 outputs are fitted to SMPL solely for presentation; metrics remain computed from their canonical joints.

Metric Contract

The table is intentionally broad: semantic correctness alone cannot establish edit quality.

Factor preservation: S-Acc and A-Acc use held-out style/action prototype retrieval. Style and Content tracks swap which factor should change and which should be retained.
Semantic alignment: MM-Dist and T2M R@3 compare generated motion with the paired target caption in Universal SMPL-H uTMR space.
Paired target: M2M R@3 and M2M-Dist evaluate retrieval and distance against the exact paired target motion.
Distribution: FID uses per-sample L2-normalized motion embeddings; Diversity is reported without converting it into a ranking objective.
Physical quality: Foot skating and Jitter are recomputed from canonical 30 fps motion, not copied from method papers.
Ranking: The displayed rank uses M2M R@3 among exact-source generated methods. Source and MotionCLR diagnostics remain visible but unranked.