A text-to-motion diffusion model can check its own turn and lateral direction: a label-free paired mirror score
Year: 2026
Type: preprint
Abstract: Preprint (not peer reviewed). Text-to-motion models often move to the wrong side: "turns left" produces a right turn. Standard evaluators show severe deficits in mirror sensitivity, so someone must check direction. We show that a motion diffusion model can check itself. For a caption c and its mirrored caption c' (left and right swapped), the paired score delta(x; c) = err(x | c') - err(x | c), the difference in the model's own denoising error at the noisiest step, ranks generated clips by whether they follow the direction word. We use it two ways: best-of-K selection, and Verify-and-Reflect, which scores one clip and mirrors it when delta < 0. No labels, retraining or external judge are used to select. On HumanML3D with MDM-50, MDM-1000 and the latent-diffusion model MLD, best-of-4 by delta beats random selection on turn and lateral actions (alt checker, pair-strict accuracy, 95% cluster-bootstrap CIs above zero), beats plain likelihood and a CompLift-style null-caption lift on the same samples, and (on MDM-50 and MLD, where we measured it) beats an external-evaluator reranker. On KIT-ML our pre-registered replication rule (turn and lateral both) is only half met: turn replicates (+0.148 [+0.053, +0.246] at K=8, after a borderline K=4 run), lateral does not. The score costs no measurable quality on MDM and about +1.5 FID on MLD. Effects are modest on the strongest model (turn +0.096 on MDM-50) and large on weaker ones (+0.29 on MLD). We state what does not work: hand direction works only on MLD, lateral direction does not transfer to KIT, and using the same signal for training (filtering flipped captions) helped only lateral on MDM-50, so the contribution is test-time.
Author: Hanzala, Muhammad
Language: English
Topic: Human Motion and Animation
Field: Computer Science
Domain: Physical Sciences
Keywords: self-checking, HumanML3D, latent diffusion models
Open Access status: green
OpenAlex ID: W7236700378
