Where Do Text-to-Motion Diffusion Models Decide Left or Right? Measuring the Decision Window and Fixing Wrong-Way Motions by Early-Commit Selection

Preprint. Text-to-motion diffusion models often move the wrong way when a prompt says "turn left", "step right" or "raise the left hand", and the usual relative pair metric hides it. We score a pair of prompts strictly: both the left and the right version must move the commanded way. On HumanML3D, MDM with 50 sampling steps gets about 0.66 of the turn pairs right, MDM with 1000 steps 0.32 to 0.44, and MLD about 0.10. Prompt switching with shared noise shows that, on MDM, the text fixes the direction within the first few denoising steps. We use this for early-commit selection: sample up to K noises, read the commanded direction from the early clean-motion prediction with a rule-based geometric checker, and keep the first noise whose prediction has the commanded sign. A rule frozen before the confirmation runs raises pair accuracy on fresh pairs for turns, sideways steps, arm side and forward versus backward walking, on MDM and on MLD, for about 12% more sampling time, and it does not reduce text-motion matching under a public evaluator. The gains shrink under independently defined evaluators, and a joint rule that asks two measures to agree narrows the gap. On MLD, which barely reads the direction word for turns and sideways steps, selection acts mostly as rejection sampling. We state the limits of the checker, the pair sets and the human evidence.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24222545
Primary Topic
Human Motion and Animation
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Where Do Text-to-Motion Diffusion Models Decide Left or Right? Measuring the Decision Window and Fixing Wrong-Way Motions by Early-Commit Selection

Muhammad Hanzala
Zenodo (CERN European Organization for Nuclear Research)
Human Motion and Animation
preprint

Where Do Text-to-Motion Diffusion Models Decide Left or Right? Measuring the Decision Window and Fixing Wrong-Way Motions by Early-Commit Selection

Muhammad Hanzala
preprint en

Abstract

Preprint. Text-to-motion diffusion models often move the wrong way when a prompt says "turn left", "step right" or "raise the left hand", and the usual relative pair metric hides it. We score a pair of prompts strictly: both the left and the right version must move the commanded way. On HumanML3D, MDM with 50 sampling steps gets about 0.66 of the turn pairs right, MDM with 1000 steps 0.32 to 0.44, and MLD about 0.10. Prompt switching with shared noise shows that, on MDM, the text fixes the direction within the first few denoising steps. We use this for early-commit selection: sample up to K noises, read the commanded direction from the early clean-motion prediction with a rule-based geometric checker, and keep the first noise whose prediction has the commanded sign. A rule frozen before the confirmation runs raises pair accuracy on fresh pairs for turns, sideways steps, arm side and forward versus backward walking, on MDM and on MLD, for about 12% more sampling time, and it does not reduce text-motion matching under a public evaluator. The gains shrink under independently defined evaluators, and a joint rule that asks two measures to agree narrows the gap. On MLD, which barely reads the direction word for turns and sideways steps, selection acts mostly as rejection sampling. We state the limits of the checker, the pair sets and the human evidence.

Zenodo (CERN European Organization for Nuclear Research)
Human Motion and Animation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon