Motion Editing · MotionFix Instructions

MotionFix official test benchmark for editing a supplied source motion with a free-form instruction. Every method is converted to the same nominal-30fps HML263 representation and evaluated against the paired target with MotionLab's released text_mot_match motion encoder.

Native Batch-32 paired-target protocolInspect motions
MotionFix official test1,013 source–instruction–target tripletsNominal 30 fpsMotionLab text_mot_match evaluatorTarget-length-sorted Batch-32 galleries
0
Measured methods
1,013
Paired test cases
R@3₃₂
Displayed rank metric
5
Six-way audit cases

Method Comparison

All generated methods are measured under the same native Batch-32 protocol; GT and source input remain unranked references.

Measured Values

Normalized Profile

100 is the best generated-method value on each axis.

Results

GT is pinned first; generated methods are ranked by Batch-32 R@3.

FID · Batch-32 R@1/2/3 · AvgR · M2M-Dist · Diversity
Best methodSecond methodGT and source input are unranked references

Qualitative Audit

All 1,013 test cases use synchronized Three.js views; case 0000 MotionLab uses the official SMPLify fit.

0.00s0.00s
Loading SMPL-H meshes…
All panels share the same template, lighting, camera and synchronized timeline. SMPLify-certified cases use optimized SMPL pose parameters rather than fast joint IK.

Protocol

A single native motion-embedding contract for every instruction editor.

Inputs
The released source motion and free-form MotionFix instruction. Methods with additional reference inputs belong to separate tracks.
Targets
The paired MotionFix target motion for all 1,013 cases; target captions are not available.
Retrieval
R@1/2/3 and AvgR use target-length-sorted candidate batches of 32, matching the released MotionLab evaluation loader.
Distribution
FID, M2M-Dist, and Diversity use the same released text_mot_match motion embeddings.
Ranking
Generated methods are formally ranked by Batch-32 R@3. GT target and the unedited source input are calibration references and never enter best/second-best selection.