A new preprint reports better benchmark results alongside faster generation and lower peak GPU memory for its text-to-motion students. In a controlled HumanML3D comparison, the TACD student recorded a FID score, the paper's motion-quality measure, of 1.041, compared with 2.486 for an uncapped student. The abstract describes the difference as a 58% reduction in eight-step student FID relative to distillation without the bound.
What the method changes
Text-to-motion systems turn written prompts into movement sequences. The study asks whether terminal-amplification control can improve few-step distillation when students learn from text prompts and pretrained teachers, without real-motion training targets. TACD supervises clean-motion predictions along student-generated trajectories and ties the latest teacher query to the student's step size to bound supervision weights. For diffusion teachers, it uses endpoint matching.
The method targets a weighting issue in velocity matching. The paper states that velocity matching is clean-motion matching with an inverse-square time-dependent weight. In ordinary language, TACD places a step-size-aware limit on that supervision at the terminal end of the schedule, using a capped teacher grid and a terminal margin. That is the control tested in the paper's few-step comparisons.
A result tied to the terminal coefficient
The main flow comparison used the 0.46B HY-Motion Lite motion backbone with frozen Qwen3-0.6B conditioning. Controls used 6,000 prompts, 8,000 updates and eight-step inference. The primary HumanML3D rows average three training seeds, while evaluator intervals, where reported, are based on 20 repetitions. Trained rows use the final iterate of a fixed budget rather than a checkpoint selected by validation FID.
In the terminal-coefficient intervention, restoring the full coefficient on the capped grid without moving the queries gave a mean FID of 2.669 across the reported seeds. The reported value was higher than the 1.041 recorded in the capped TACD comparison.
The comparison extended beyond one benchmark
The pattern was reported with other teacher types as well. On HumanML3D, four-step TACD with the MLD teacher had FID 0.222 versus 0.387 and R@3, a text-motion retrieval measure, of 0.818 versus 0.795. With MotionLCM-V2, TACD FID was 0.055 versus 0.066, while R@3 was 0.824.
On KIT-ML, TACD FID was 1.038 at two steps, 0.852 at four, and 0.649 at eight, compared with 1.530, 0.976 and 0.745 for the truncated teacher. The paper reports reductions of 32% at two steps and 13% at four and eight steps. On Kimodo at eight steps, TACD with the 8B encoder had R@1 of 71.39 and FID of 0.054, versus 69.36 and 0.064 for the truncated teacher. The 1B-encoder version had R@3 of 86.66 and FID 0.054, with peak memory of 4.0 GB.
The gains come with a narrower claim
Kimodo requires a separate reading of the FID figures. The paper calculates Kimodo FID in TMR joint space without real motions, and says it is not comparable with HumanML3D FID. A lower number on one benchmark therefore cannot be treated as a direct numerical comparison with the other.
The efficiency gap was also large under the paper's reported RTX L40 protocol. Eight-step HY-Motion students were 7.7 times faster end to end and used 6.7 times less peak memory than their teachers. On Kimodo, students were 11.9 times faster and used 3.8 times less peak memory. These figures describe the reported protocol and hardware measurement.
This remains a preprint, identified as arXiv:2610.02867v1 and dated 2 Oct 2026. Its main HumanML3D rows average three training seeds, with evaluator intervals based on 20 repetitions, and the trained rows use a fixed-budget final iterate rather than validation-FID checkpoint selection. The study's reported comparisons use that evaluation setup, and Kimodo and HumanML3D FID values must be read on separate scales.
Paper data and sources
Original title: TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control
Authors: Wei-Jin Huang, Yuan-Ming Li, Kun-Yu Lin et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-02
DOI: Not available
Original paper · Full text