acceptodds
Under review as a conference paper at ICLR 2027

TACD: Distilling Efficient Text-to-Motion Models via Terminal Amplification Control

Abstract

Recent text-to-motion models have improved motion quality and instruction following, yet many-step denoising and large model components make deployment slow and memory-intensive. We present Terminal-Amplification-Controlled Distillation (TACD), an on-policy approach for training efficient motion generators from text prompts and pretrained teachers, without real-motion training data. Building on segmented on-policy flow distillation, we supervise clean-motion predictions along student-generated trajectories. We identify a failure mode in which velocity matching on a fixed supervision grid repeatedly overweights errors near the denoising endpoint, degrading few-step generation. TACD ties the latest teacher query to the student’s step size, bounding the effective loss weights in clean-motion space without changing inference. Experiments on HumanML3D and KIT-ML demonstrate improved few-step generation, including a 58% reduction in eight-step HY-Motion student FID relative to distillation without this bound. For diffusion teachers, the endpoint-matching form of TACD yields four-step students with lower FID and matched or improved text–motion retrieval relative to their 50-step teachers on HumanML3D. On HY-Motion and Kimodo, eight-step students with compact components achieve 7.7–11.9× end-to-end speedups and reduce peak GPU memory by 3.8–6.7× relative to their teachers.

Primary area:
generative models
Keywords:
text-to-motion generation, flow matching, few-step generation, on-policy distillation, data-free knowledge distillation, loss reweighting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CQSFOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

TACD introduces a data-free, few-step text-to-motion distillation by controlling persistent near-terminal supervision amplification on fixed grids. Based on π-Flow’s single-knot version, the authors find that the inverse-square weights on clean-motion prediction errors induced by velocity matching will degrade training, especially when querying near the clean sample. To address this, TACD moves teacher queries away from these endpoints using a manual margin. Validation includes coefficient ablation, gradient analysis, various T2M quality metrics on different datasets, and deployment measurements with two large-scale text-to-motion generative models, i.e., HY-Motion and Kimodo.

Critical strengths

This work is easy to understand.

Critical weaknesses

Limited methodological novelty of endpoint amplification control.: The inverse-square amplification arising from clean-sample prediction in velocity matching, as well as its mitigation through loss-weight clipping, has been explicitly recognized and addressed in prior work. For example:

  • FrameFlow [1, Sec. 2.2] explicitly identifies exploding losses caused by the (1−t)−2(1-t)^{-2} factor and addresses them through a simple clipping strategy.
  • MultiFlow [2, Code] adopts the same stabilization strategy.
  • Phased dmd [3, Eq. 13] addresses a related inverse-square singularity in subinterval score matching through coefficient clipping.
  • JiT [4, Sec. 4.3] similarly clips the denominator when converting clean predictions into velocities.

These examples suggest that endpoint amplification and its mitigation are established considerations in generative modeling, often handled as part of the training recipe rather than as standalone methodological contributions. Furthermore, the paper's own ablations (Tab.5) show that conventional weight capping and direct endpoint matching achieve comparable performance to TACD's proposed grid modification. It remains unclear what substantive methodological advance the proposed step-size-dependent rule offers beyond these existing stabilization techniques.

The reported failure may be largely induced by the authors' choice of supervision grid. In particular, the uncapped baseline deterministically queries a near-clean endpoint at every iteration, whereas the original π-Flow implementation uses randomized intermediate queries and includes a nonzero micro-window to avoid near-clean queries in its released training configurations, which also restricts the query range.


Reference: [1] Yim, Jason, et al. "Fast protein backbone generation with se (3) flow matching." arXiv preprint arXiv:2310.05297 (2023). [2] Campbell, Andrew, et al. "Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design." ICML 2024. [3] Fan, Xiangyu, et al. "Phased dmd: Few-step distribution matching distillation via score matching within subintervals." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026. [4] Li, Tianhong, and Kaiming He. "Back to basics: Let denoising generative models denoise." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Questions for authors

What is the advantage of the proposed step-size-dependent bound over existing stabilization techniques? Prior work has already addressed inverse-square amplification through denominator or loss-weight clipping. Moreover, Table 5 shows that conventional weight capping (FID 1.011) and direct endpoint matching (FID 1.029) perform comparably to TACD's grid modification (FID 1.041).

6 comments
AuthorsOpenReview ↗

Response to Reviewer CQSF (1/3): what a step-tied bound adds over existing stabilizers (Q, W1)

Thank you for the review and the precise pointers to prior work. The inverse-square factor and its clipping are established in prior work. TACD's contribution is to tie terminal-weight control to the student's step size and to examine its consequences for few-step on-policy distillation. Table 5's controlled interventions test whether terminal weighting drives the degradation; the new cross-step experiments and the derivation (here and in general response 2/2) address how to choose the bound, and comment 3/3 applies the same rule to three more flow teachers.

Table 5. Its weight-capping row is conventional coefficient clipping, set at the step-tied level C = 4K² = 256 (Sec. 4): TACD's bound applied to the coefficient rather than to the grid. Table 5 therefore compares three ways to stay under one bound: grid cap 1.041, endpoint matching 1.029, coefficient at 4K² 1.011, against 2.486 unbounded. Restoring C = 10⁶ on the capped grid, without moving any query, returns FID to 2.669. TACD's rule is the step-tied bound b ≤ 4K² (Sec. 3.3); moving the query is how the paper applies it to flow teachers.

Prior constants. Written as bounds on b = (1 − s)⁻², the stabilizers you cite use fixed constants:

stabilizerwhat it clipsb ≤
Phased DMD (Eq. 13)its 1/σ² coefficient, to [0, 10], "to prevent overflow"10
FrameFlow, MultiFlowt at 0.9100
JiTt_eps = 0.05400
π-Flow's released windowqueries kept 3/128 from the clean end1,820 (our time axis)

None is tied to the sampler's step count.

The failure they would need to catch is not an overflow. The unbounded student is numerically stable. Unit-norm gradient clipping binds on at least 799 of 800 logged updates, the median pre-clip norm is 245-263 (5.0-6.1 for the bounded students), and no update overflows. FID is still 2.4× worse (2.486 vs 1.041), and at the end of training 96.6% of its gradient norm sits on the final segment. An overflow guard does not see this failure, so it gives no reason to pick a particular bound.

What the step-tied bound adds: one rule across step counts. Same student, teacher, grid and budget; FID ↓; ± = SD over three seeds; † = seed 42 only (at K = 2 its % is against the grid rule's seed 42, 9.73); other cells are the paper's three-seed means:

bound on the terminal weight bb ≤K = 8K = 2
at or below 4K² (TACD)
grid rule δ = 1/(2K) (paper default)4K²1.04 ± .039.37 ± .32
coefficient cap at 4K² (Table 5, weight capping)4K²1.019.89†
endpoint matching (Table 5)11.039.11†
fixed constants
clip at 100 (FrameFlow, MultiFlow)1001.01†10.76 ± .12 (+15%, every seed)
clip at 400 (JiT)400–12.09 ± .17 (+29%, every seed)
margin δ = 3/32 (Table 18)1141.03†9.49 ± .12 (tie)
margin δ = 1/32 (Table 18)1,0241.00†12.38 ± .22 (+32%, every seed)
cap at 4,096 (Fig. 6)4,0961.07†15.39† (+58%)
no bound≈10⁶2.4916.44†

At eight steps every fixed constant we ran on the grid lands at 1.00-1.07, so an eight-step sweep cannot tell a safe constant from an unsafe one. At two steps they separate: the clips at 100 and 400 and δ = 1/32, the best margin at eight steps, are worse on every seed and on all five metrics, and the cap at 4,096 is back near the unbounded student. The clip-100 student's gradient norms match ours (median 7.4 vs 7.2), so nothing in training flags it. Every bound at or below 4K² stays in the best group at both step counts; the constants that fail are too loose for the step count.

AuthorsOpenReview ↗

Response to Reviewer CQSF (2/3): why 4K², and the supervision grid (W2)

Why 4K². In the student's update (Sec. 3.1), step k uses the velocity (x̂₁ − x_k)/(1 − t_k), so an error in its clean-motion prediction enters that velocity with squared weight (1 − t_k)⁻². That is the same form as the coefficient b = (1 − s)⁻² that velocity matching gives a query at s (Sec. 3.2). On the uniform grid, every query before the final step has b at most 4 times its own step's weight; the worst case is the end of the second-to-last step. The midpoint rule gives the final step the same factor at every K: b ≤ 4K² = 4(1 − t_{K−1})⁻². A fixed bound C gives the final step a factor of C/K² instead, which grows 16-fold from eight steps to two (the clip at 100: 1.6 at K = 8, 25 at K = 2). This bounds the coefficient ratio, not quality directly. In the comparison of comment 1/3, every rule within the factor 4 stays in the best group; beyond it, results vary.

W2: the supervision grid and π-Flow's micro-window. We agree that π-Flow's released configs keep every query at least 3/128 from the clean end. For the single-knot DX policy the window average equals the instantaneous velocity (App. B), so the window's only effect is a terminal margin, the mechanism we identify. Its bound (b ≤ 1,820) is fixed, while 4K² is 256 at eight steps and 16 at two. We ran π-Flow's own schedules with the paper's student, teacher calls and budget. FID ↓, mean ± SD over three seeds; "every seed" means the paired difference has the same sign on all three seeds:

queriesK = 8K = 4K = 2
uniform.765 ± .0151.168 ± .0574.77 ± .23
uniform + TACD.755 ± .032 (tie)1.080 ± .046 (every seed)3.53 ± .05 (every seed)
stratified.742 ± .0171.086 ± .0434.62 ± .23
stratified + TACD.700 ± .009 (every seed)1.003 ± .067 (every seed)3.43 ± .08 (every seed)

Random queries without a window still put a query with b > 4K² into 93% (uniform) or 100% (stratified) of updates, up to 10⁶. π-Flow's released recipe (random queries, 3/128 window) at two steps, with R@1, R@3 and MM-Dist as three-seed means that improve on every seed:

π-Flow's released recipe, K = 2FID ↓ s42 / s43 / s44mean ± SDR@1 ↑R@3 ↑MM-Dist ↓
released window3.81 / 3.50 / 3.563.62 ± .16.395.6973.77
final segment bounded by 4K²3.26 / 3.23 / 3.173.22 ± .04.405.7093.69

At two steps, bounding the released window's final segment by 4K² lowers FID by 11% on every seed; at eight steps stratified + TACD (.700 ± .009) is our best result. With π-Flow's uniform and stratified queries (first table), the bound lowers FID on every seed in five of the six cells and ties in the sixth, and the gradients are steadier: the 99th-percentile norm is about 3× lower at eight steps and 8-12× lower at two (general response 1/2).

So the deterministic grid makes the failure extreme (2.49 vs 1.04 at eight steps). The failure itself, and the bound's FID gain, remain under π-Flow's random queries at four and two steps, and grow as the step count drops.

In the revision we will:

  • add these tables to the appendix;
  • add the derivation of 4K² to Sec. 3.3;
  • add FrameFlow, MultiFlow and Phased DMD to Sec. 2, next to JiT and Hong et al. (2026), which the paper already cites for the inverse-square factor and capped weights;
  • update the contrast between the grid and random sampling (Secs. 1-3) with these statistics.
AuthorsOpenReview ↗

Response to Reviewer CQSF (3/3): the same rule on three more flow teachers (W1)

Is the bound a recipe detail of one setup? W1 notes that terminal-weight control is usually handled inside a training recipe. To check that the step-tied rule is not specific to HY-Motion, we applied it unchanged to three more flow teachers of different designs. The rule has no constant to tune: each student's final-segment query is capped at its midpoint, so b ≤ 4K² for that student's own K. Each student was distilled for 8k updates with three seeds and scored with its teacher's own evaluator:

  • MotionHiFlow: a latent DiT that samples in three temporal stages (14 velocity calls); the bound is applied at the end of each stage.
  • MoGeFlow-L: a flow over part-level code embeddings; the bound is applied before its final snap to codes.
  • MARDM: we distil only the flow head inside each of its 18 masked-autoregressive steps.

FID; "teacher" is the teacher itself sampled with that many steps; students are three-seed means:

modelstepsteacherno boundTACD
MotionHiFlow14 calls (default).032
6.043.042.034
32.031.89.045
MoGeFlow-L96 (default, our run).071
8.630.190.090
41.451.51.143
MARDM (flow head)adaptive (default).125
8.191.129.144
4.230.192.143

Every difference between the two student columns has the same sign on all three seeds, and at every step count the TACD student beats the teacher sampled with the same number of steps. As on HY-Motion, every unbounded run stays finite and completes all 8k updates under the usual gradient clipping; the failure shows only in sample quality. Without the bound, MotionHiFlow's 3-step student stays close to the teacher's own 3-step result, and MoGeFlow-L's 4-step student is ten times worse than with it. With the bound, MotionHiFlow's 6-step student is near the 14-call default, and MoGeFlow-L's 8-step student is within .02 of its 96-step default. On each of these teachers the gain is largest at the fewer steps, as on HY-Motion. On MARDM at eight steps, where the unbounded head is already close to the teacher's adaptive default, the bound costs .015.

Reviewer_dLyCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Although modern text-to-action models can generate high-quality motion trajectories, they suffer from slow multi-step reasoning and high memory requirements. To address this issue, the authors propose a method called Terminal-Augmented Control Distillation (TACD). It is a data-free, policy-based distillation technique used to train fast student generators. The authors found that standard velocity matching on a fixed grid over-penalizes terminal denoising errors, thereby reducing the quality of multi-step synthesis. Therefore, TACD effectively mitigates this issue by aligning the latest teacher query with the student’s step to constrain the clean motion loss weights. On the HumanML3D and KIT-ML datasets, TACD reduced the FID of HY-Motion by 58% using only eight steps. Furthermore, the four-step student outperforms the 50-year diffusion teacher, achieving a 7.7- to 11.9-fold increase in inference speed and a 3.8- to 6.7-fold reduction in memory usage.

Critical strengths
  • This paper presents a comprehensive causal analysis of velocity matching.
  • This paper demonstrates strong experimental results. Across five teachers and three benchmark tests, it achieved a 58% reduction in the FID at level 8, outperforming the level-50 diffusion teacher by 4 levels.
Critical weaknesses
  1. This paper focuses on text-to-motion generation, but its innovation lies in the OPD design rather than the motion characteristics. Specifically, the inverse-square weighting mechanism does not involve any motion-specific features, and the authors chose a fixed grid arbitrarily rather than having it determined by task constraints. This results in a significant disconnect between the paper’s positioning and its actual contributions.
  2. Missing baselines. This paper uses only the baseline DX strategy while ignoring the more performant GMFlow. Consequently, the contribution to π-Flow appears limited. Furthermore, the paper fails to compare its approach with π-Flow’s random query sampling, which only occasionally produces large weights. These issues mean that the necessity of the proposed remedy has not yet been substantiated.
  3. The evaluation overlooks the natural benchmark of distilling pre-trained models on public datasets for specific tasks—such as initializing trajectories from forward-diffusion-based HumanML3D actions—which similarly addresses the issue of missing training data. Furthermore, since the MLD, MDM, and MotionLCM-V2 teachers have already been trained on HumanML3D and KIT-ML, adopting a completely data-free setup may not be appropriate.
Questions for authors

See weaknesses. If my concerns are adequately addressed, I would consider increasing my score.

8 comments
AuthorsOpenReview ↗

Response to Reviewer dLyC (1/3): π-Flow's random queries and GM policy (W2)

Thank you for the review and for saying what would change your assessment. We ran new experiments for each point. This comment covers π-Flow's random and stratified queries and its GM policy (W2); comment 2/3 covers three more flow teachers (W2); comment 3/3 covers motion-level metrics (W1) and data-seeded rollouts (W3).

Random queries (W2, third point). Same student, teacher calls and 8k updates as the paper, three seeds. Queries are drawn uniformly within each segment (π-Flow, Alg. 1) or stratified, with and without the step-tied bound. FID, mean ± SD; "every seed" means the paired difference has the same sign on all three seeds.

queriesK = 8K = 4K = 2
uniform.765 ± .0151.168 ± .0574.77 ± .23
uniform + TACD.755 ± .032 (tie)1.080 ± .046 (every seed)3.53 ± .05 (every seed)
stratified.742 ± .0171.086 ± .0434.62 ± .23
stratified + TACD.700 ± .009 (every seed)1.003 ± .067 (every seed)3.43 ± .08 (every seed)

Each update places four queries in the final segment, so most updates contain one with b = (1 − s)⁻² > 4K², up to 10⁶. Pre-clip gradient norm (training clips at 1) over all 8k updates, ranges over seeds:

queriesupdates with b > 4K² (K = 8)P99, K = 8P99, K = 2
uniform93%249-2561,058-1,490
uniform + TACD076-88129-136
stratified100%231-2471,224-1,479
stratified + TACD073-78133-149

Random queries soften the failure at eight steps (a tie with uniform queries, −6% on every seed with stratified ones), but they do not remove it at four steps (−7 to −8% on average, every seed) or two (−26%).

π-Flow's released recipe, with its 3/128 micro-window, also leaves a gap at two steps. Bounding only its final segment by 4K² improves FID on every seed, and R@1, R@3 and MM-Dist with it (three-seed means):

π-Flow's released recipe, K = 2FID ↓ s42 / s43 / s44mean ± SDR@1 ↑R@3 ↑MM-Dist ↓
released window3.81 / 3.50 / 3.563.62 ± .16.395.6973.77
final segment bounded by 4K²3.26 / 3.23 / 3.173.22 ± .04.405.7093.69

The deterministic grid makes the failure extreme (2.49 vs 1.04 at eight steps), but it does not create it. Stratified + TACD (.700) is our best eight-step result and the recipe we recommend.

GM policy (W2, first point). The failure is not specific to π-Flow: any student that predicts a clean sample and is trained through a velocity target carries this inverse-square factor, which is why FrameFlow, MultiFlow and JiT clip it (comment 2/3). π-Flow's GM policy changes what the student outputs: with positive mixture variance, the loss gradient with respect to the GM means vanishes as s → 1, so the inverse-square argument does not carry over to it. Measured directly, GM trails DX with the bound in six of the seven settings below and ties it in the seventh. We ran GM (ported from π-Flow's released code, eight components; per stage on MotionHiFlow) without a bound, same budget, three seeds, on HY-Motion (stratified queries), MotionHiFlow and MoGeFlow-L (comment 2/3). FID:

DX, no boundGM, no boundDX + TACD
HY-Motion, 2 steps4.625.283.43
HY-Motion, 4 steps1.0861.0601.003
HY-Motion, 8 steps.742.697.700
MotionHiFlow, 3 steps1.892.94.045
MotionHiFlow, 6 steps.042.155.034
MoGeFlow-L, 4 steps1.51.215.143
MoGeFlow-L, 8 steps.190.092.090

On HY-Motion at eight steps GM and DX with TACD tie (.697 vs .700); in every other row, MoGeFlow-L at eight steps included, DX with TACD is better on every seed.

AuthorsOpenReview ↗

Response to Reviewer dLyC (2/3): three more flow teachers and existing stabilizers (W2)

Generality across settings (W2, second point). Our contribution is to characterize a recurring failure mechanism in few-step distillation and to control it with a rule tied to the student's step count, with controlled evidence across query schedules and teacher architectures. The failure is silent: the unbounded student never overflows, its updates are clipped as usual, and its FID is still 2.4× worse on the paper's eight-step grid. It also extends beyond the paper's deterministic grid and beyond HY-Motion: comment 1/3 shows it under π-Flow's random queries, and here we distilled three more flow teachers of different designs, each for 8k updates with three seeds and scored with its own evaluator: MotionHiFlow (a latent DiT that samples in three temporal stages, 14 velocity calls; bound at the end of each stage), MoGeFlow-L (a flow over part-level code embeddings; bound before its final snap to codes) and MARDM, where we distil only the flow head inside each of its 18 masked-autoregressive steps. FID; "teacher" is the teacher itself sampled with that many steps; students are three-seed means:

modelstepsteacherno boundTACD
MotionHiFlow14 calls (default).032
6.043.042.034
32.031.89.045
MoGeFlow-L96 (default, our run).071
8.630.190.090
41.451.51.143
MARDM (flow head)adaptive (default).125
8.191.129.144
4.230.192.143

Every difference between the two student columns has the same sign on all three seeds, and at every step count the TACD student beats the teacher sampled with the same number of steps. Unbounded, MotionHiFlow's 3-step student stays close to the teacher's own 3-step result; at 6 steps TACD comes near the 14-call default and also improves R@1, R@3, MM-Dist and Diversity. On MARDM at eight steps, where the unbounded head is already close to the teacher's adaptive default, the bound costs .015. The stabilizers we know of (FrameFlow, MultiFlow, JiT, Phased DMD, π-Flow's window) use fixed constants. General response (2/2) shows that on the same grid the FrameFlow/MultiFlow clip at 100 ties at eight steps but loses 15% at two, and JiT's clip at 400 loses 29% at two, each loss on every seed, while every bound at or below 4K² in that table stays in the best group without tuning. It also derives 4K² from the student's own step weights.

AuthorsOpenReview ↗

Response to Reviewer dLyC (3/3): motion-specific positioning (W1) and data-seeded rollouts (W3)

W1: positioning. We agree that the weighting mechanism itself is general. This paper studies its effect on few-step text-to-motion distillation and introduces TACD, which ties the bound on this weight to the student's step count. Our motivation is to obtain efficient motion generators from released teachers and text prompts, without access to the teachers' motion training data. We will make the distinction between the general mechanism and the motion setting explicit in the introduction.

We also agree that the fixed supervision grid is our training choice, not a task constraint; it makes the terminal amplification easy to isolate. Comment 1/3 shows that, at few steps, neither the problem nor the fix depends on it: with π-Flow's random and stratified queries the bound improves FID at four and two steps on every seed, and with π-Flow's released micro-window at two steps.

The weighting also has visible consequences for the generated motion. In the controlled HY-Motion experiment on HumanML3D (K = 8, three seeds, 4,648 motions per arm), the unbounded student's motion is damped. Power above 3 Hz, as a fraction of real motion's (real = 1):

no boundTACD50-step teacher
feet.55.84.71
hands.30.56.46

TACD is closer to real motion here than the 50-step teacher. Across 15 physical measures (foot sliding, floating, foot height and penetration; jerk and power above 3 Hz of feet, hands and root; root and joint speed; contact rate), TACD is closer to real motion than the unbounded student on 10 of them. These results tie the weighting analysis to concrete motion quality, beyond FID and retrieval. We will add the full table to the appendix.

W3: data-seeded rollouts. We ran the baseline you suggest on MLD (K = 4, three seeds, 30k updates). MLD is a favourable case for it, because HumanML3D is MLD's own training data. A fraction of each batch starts its segments from forward-diffused HumanML3D training motions instead of the student's own rollout; everything else is unchanged.

rollout startsFID ↓R@1 ↑R@3 ↑MM-Dist ↓
student's own (data-free, paper).222 ± .002.525.8182.909
50% data-seeded.222 ± .001.523.8172.915
100% data-seeded.219 ± .001.522.8142.926

Seeding with real motions gives no consistent improvement across quality and retrieval metrics. At 100% FID drops by .003, while R@1, R@3 and MM-Dist get slightly worse on every seed; at 50% FID is unchanged and the other three are again slightly worse on every seed. MotionLCM-V2 distils with forward-diffused HumanML3D motions. In our evaluation it reaches .072, against .055 for data-free TACD at the same architecture and step count (Table 9).

Data-free distillation is a capability: it distils a released teacher into a few-step student using only text prompts and the teacher's weights, which is all that a release such as HY-Motion provides; no real-motion training examples are needed. The public-data teachers (MLD, MDM, MotionLCM-V2) are controlled testbeds for checking whether motion data would help, including against rollouts seeded with their own training motions. On MLD, even the teacher's own training motions give no consistent gain over the data-free protocol. We will add the data-seeded rows.

Reviewer_PHxAOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Terminal-Amplification-Controlled Distillation (TACD) for efficient text-to-motion generation without real-motion training targets. The method builds on segmented on-policy distillation, where a student predicts the clean motion at each segment start and receives teacher supervision along its own detached trajectory. The paper observes that velocity matching is equivalent to clean-motion matching weighted by (1−s)−2(1-s)^{-2}, so a fixed query near the endpoint receives a very large coefficient at every update. TACD bounds this coefficient by tying the latest teacher query to the student step size. For diffusion teachers, it uses direct endpoint matching. Experiments cover HY-Motion, MLD, MotionLCM-V2, MDM, and Kimodo teachers on HumanML3D, KIT-ML, and Kimodo, together with semantic, human-preference, and efficiency evaluations.

Critical strengths

The terminal-amplification claim is supported by careful controlled evidence. The paper derives the inverse-square coefficient, measures its gradient concentration, and intervenes on the supervision time and coefficient separately. Grid capping, endpoint matching, and weight capping all reduce the eight-step HY-Motion student FID from 2.486 to about 1.01–1.04, while restoring the 10610^6 coefficient raises it to 2.669. The coefficient sweep, last-point controls, two student sizes, and three training seeds make the causal interpretation substantially more convincing than a standard component ablation.

The evaluation is broad and practically relevant. The paper evaluates multiple teachers and three motion benchmarks, repeats the HumanML3D comparison with a second evaluator, and includes a VLM judge, a blind human study, qualitative results, and measured latency and memory. The eight-step students provide 7.7–11.9$\times$ end-to-end speedups, and the paper separates the effects of fewer denoising steps and smaller conditioning encoders. These results show that the proposed supervision design can support useful few-step motion generators.

Critical weaknesses

The main baseline does not follow the original π\pi-Flow query schedule. Appendix B states that π\pi-Flow samples query times uniformly, whereas the uncapped student uses the authors’ deterministic grid with a query fixed at s=.999s=.999 in every update. TACD therefore fixes a severe weighting pathology of this fixed-grid variant, but the paper does not compare with randomized or stratified within-segment queries under the same student, number of teacher calls, and training budget. Table 18 further shows that less extreme margins, δ=1/32\delta=1/32 and 3/323/32, already obtain FID 0.996 and 1.033, comparable to TACD’s 1.041, the headline 58% reduction is therefore relative to a near-singular δ=10−3\delta=10^{-3} baseline rather than a reasonable alternative schedule. Direct endpoint matching and coefficient capping also perform as well as the proposed grid rule, while prediction–loss alignment and capped weighting are already discussed in prior work. The evidence supports controlling the coefficient, but does not establish the necessity or novelty of TACD’s particular construction.

The generalization and efficiency positioning remain limited. Grid-capped flow TACD is evaluated only with HY-Motion on HumanML3D, using uniform grids at four and eight steps. The other teacher and dataset experiments instead use endpoint matching, so they do not establish that the proposed step-size-dependent flow bound transfers to other flow settings. The reported 7.7–11.9$\times$ speedups are also measured only against the corresponding 50- or 100-step teachers, without a matched comparison to recent efficient text-to-motion generators. Since published latency numbers use different hardware, sequence lengths, and text-encoder accounting, a common protocol is needed to establish the absolute quality–efficiency trade-off. The Kimodo training captions and their overlap with the 607 evaluation prompts are also unclear.

Questions for authors
  1. Could you compare TACD with the original π\pi-Flow strategy that samples query times randomly, as well as a stratified-random variant, using the same student, number of teacher queries, and training iterations? Please report seed-level quality and gradient statistics. This would clarify whether TACD addresses a general on-policy distillation problem or a pathology introduced by the deterministic grid. Were the DMD, consistency, and Reflow ports also trained to convergence under their shared 8k-update budget?
  2. Can the authors evaluate grid-capped flow TACD on another flow teacher or dataset, or at least at two steps and with a nonuniform sampling schedule? For the diffusion experiments, could you compare endpoint matching with noise- or velocity-space matching on the same on-policy DDIM states? This would distinguish transfer of terminal-amplification control from the general benefit of training directly in clean-motion space.
  3. Which prompts and splits are used to train the Kimodo students, and are any of the 607 evaluation captions observed during distillation? Please report results on a strictly unseen-caption subset. Could you also report the training cost, discuss how TACD differs from recent efficient generators [1–5], and compare its latency–quality trade-off with one or two representative methods under the same hardware and output length?

[1] Dai, Wenxun, et al. "Motionlcm: Real-time controllable motion generation via latent consistency model." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[2] Zhou, Wenyang, et al. "Emdm: Efficient motion diffusion model for fast and high-quality motion generation." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[3] Zhang, Zhiyuan, and Lingqiao Liu. "MoSCo: Real-time and Efficient Text-to-Motion Synthesis via Delta Training." 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026.

[4] Fu, Dongjie, et al. "Mogo: Residual quantized hierarchical causal transformer for real-time and infinite-length 3d human motion generation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. No. 5. 2026.

[5] Zheng, Zhiwei, et al. “Next-Scale Autoregressive Models for Text-to-Motion Generation.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 16376–86.

8 comments
AuthorsOpenReview ↗

Response to Reviewer PHxA (1/5): π-Flow's query schedules (Q1)

Thank you for the careful review and the concrete requests. We ran each of them. This comment covers Q1. Comment 2/5 covers W1; 3/5 the ports, other grids and flow teachers; 4/5 the Kimodo captions, training cost and efficient generators; 5/5 the diffusion matching spaces.

Setup (all new Lite runs): the paper's HY-Motion Lite student and teacher, data-free prompts, the same number of teacher calls per update, 8k updates, the paper's evaluation, three training seeds (42-44). "Every seed" means the paired difference has the same sign on all three seeds. TACD caps only the final segment's queries at its midpoint, 1 − Δt_last/2 (1 − 1/(2K) on uniform grids).

Q1: π-Flow's random and stratified queries, K = 8. Gradient statistics are the pre-clip norm (training clips at 1) over all 8k updates.

queriesFID s42 / s43 / s44mean ± SDgrad. norm median / P99updates with b > 4K²max b
uniform (π-Flow Alg. 1).752 / .781 / .761.765 ± .0158.8-8.9 / 249-25693%≈10⁶
uniform + TACD.780 / .766 / .720.755 ± .0327.4-7.7 / 76-880256
stratified.761 / .728 / .738.742 ± .0178.6-8.7 / 231-247100%≈10⁶
stratified + TACD.703 / .708 / .690.700 ± .0097.3-7.5 / 73-780256
paper's grid2.56 / 2.42 / 2.482.49 ± .07245-263 / –100%≈10⁶
paper's grid + TACD1.01 / 1.07 / 1.041.04 ± .035.0-6.1 / –0256

At four and two steps (FID; TACD is better on every seed in all four comparisons):

queriesK = 4K = 2K = 2 grad. P99
uniform1.168 ± .0574.77 ± .231,058-1,490
uniform + TACD1.080 ± .0463.53 ± .05129-136
stratified1.086 ± .0434.62 ± .231,224-1,479
stratified + TACD1.003 ± .0673.43 ± .08133-149

At two steps R@1, R@3 and MM-Dist also improve on every seed for both schedules.

Reading: a general problem that the grid makes extreme. Random queries do not avoid the near-clean coefficient. Each update places four queries in the final segment (two per segment, two micro-batches), and any query in its upper half has b > 4K². So 93% (uniform) and 100% (stratified) of updates contain one. The large coefficient arrives at varying times instead of the same time every update. At eight steps the bound ties with uniform queries and, with stratified ones, lowers FID by about 6% on every seed; at four steps it lowers FID by 7-8% on average and at two by 26%, on every seed for both schedules. So with both schedules the bound is a robust addition: it lowers FID on every seed in five of the six cells and ties in the sixth, and the gradients are steadier, with the 99th-percentile norm about 3× lower at eight steps and 8-12× lower at two. Stratified + TACD (.700) is our best eight-step result and the recipe we recommend; the grid is the controlled setting.

AuthorsOpenReview ↗

Response to Reviewer PHxA (2/5): fixed margins, endpoint matching and coefficient capping (W1)

W1: fixed margins (Table 18). We agree that at eight steps the 58% is measured against the unbounded grid, and that δ = 1/32 and δ = 3/32 match TACD there. They separate at two steps. Same student, teacher, grid and budget; FID ↓; ± = SD over three seeds; † = seed 42 only; other K = 8 cells are the paper's three-seed means:

boundb ≤K = 8K = 2
margin δ = 1/32 (Table 18)1,024.996†12.38 ± .22
margin δ = 3/32 (Table 18)1141.033†9.49 ± .12
grid rule δ = 1/(2K) (TACD)4K²1.0419.37 ± .32 (s42: 9.73)
coefficient cap at 4K² (Table 5)4K²1.0119.89†
endpoint matching (Table 5)11.0299.11†
no bound≈10⁶2.48616.44†

δ = 1/32 is worse than the step-tied rule on every seed and on all five metrics; δ = 3/32 ties. The margin that looks best at eight steps is the worse one at two, and an eight-step sweep cannot tell them apart. δ = 1/(2K) uses the same rule at both step counts. General response (2/2) adds the clip at 100 (FrameFlow, MultiFlow), the clip at 400 (JiT) and π-Flow's window, and a short derivation of why the bound scales as 4K².

W1: endpoint matching, coefficient capping, prior work. Table 5's three rows are three ways to stay under one bound. The weight-capping row sets the terminal coefficient to C = 4K² = 256, which is TACD's bound applied to the coefficient; endpoint matching has b = 1. TACD's rule is the step-tied bound b ≤ 4K² (Sec. 3.3); moving the final query is how the paper applies it to flow teachers. All three stay in the best group at two steps (table above), while the clips at 100 and 400 fall out of it (general response 2/2).

The inverse-square factor, prediction–loss alignment and capped weights are prior work (Hong et al., 2026, cited in Sec. 2; FrameFlow, MultiFlow, JiT, Phased DMD), and we do not claim them. The paper adds when the bound is needed and a size that needs no tuning across step counts. On the student's own trajectory the failure is silent: nothing overflows, and FID is still 2.4× worse. On the grid, the range of safe constants also shrinks as the step count drops, while 4K² stays inside it at every step count we ran. We will add the two-step comparison with the fixed margins.

AuthorsOpenReview ↗

Response to Reviewer PHxA (3/5): the ports (Q1b), other grids and flow teachers (Q2a, W2)

Q1b: training budget and convergence. The paper compares methods at a matched update budget (8k); no method, TACD included, had converged at 8k. We retrained the three ports from scratch at twice the budget (16k updates, cosine schedule stretched, three seeds), trained TACD the same way, and ran consistency distillation (CD) at ten times the budget (80k). Its updates are cheaper, so 80k CD updates cost about as many GPU-hours as TACD's 8k (comment 4/5). FID per seed (42 / 43 / 44), mean in parentheses:

method8k updates16k updates80k updates
DMD1.77 / 1.84 / 1.88 (1.83)1.17 / 1.29 / 1.32 (1.26)–
Reflow3.41 / 3.17 / 3.44 (3.34)2.23 / 2.08 / 2.14 (2.15)–
CD3.02 / 2.85 / 2.85 (2.91)2.81 / 2.63 / 2.81 (2.75)2.70 / 2.53 / 2.65 (2.63)
TACD1.01 / 1.07 / 1.04 (1.04).89 / .94 / .95 (.93)–

CD gains little from longer training: ten times the updates gives 10%. None of these longer-trained runs reaches TACD at 8k, and at the same 16k budget TACD stays ahead on every seed. The closest, DMD at 16k, is worse on every seed and also on R@1, R@3, Diversity and MM-Dist, and took 9.9 GPU-h per run, against TACD's 5.75 at 8k. With π-Flow's stratified queries, TACD needs only 4k updates to beat every port at 16k, and it keeps improving with more (FID, seed 42, cosine schedule scaled to each budget):

updates4k8k16k32k
stratified + TACD.82.70.61.58

We will add these tables to the appendix.

Q2a / W2: two steps, nonuniform grids, another flow teacher. HY-Motion first, then three more flow teachers. Within HY-Motion, the grid rule transfers to fewer steps and to nonuniform schedules. On a nonuniform grid the rule caps the final segment's queries at its midpoint, 1 − Δt_last/2, and leaves the other segments unchanged. FID; ± = SD over three seeds; † = seed 42 only; other cells are three-seed means:

settingno boundTACD
uniform grid, K = 216.44†9.73† (three seeds: 9.37 ± .32)
nonuniform grid 0 / .4 / 1, K = 213.47 ± .206.28 ± .17
time-shifted grid 0 / .094 / .238 / .484 / 1 (shift 3.2), K = 45.14 ± .061.25 ± .05
power grid 0 / .350 / .646 / .875 / 1 (p = 1.5), K = 412.85 ± .333.28 ± .24
K = 2, π-Flow's uniform queries4.773.53 (every seed)
K = 2, π-Flow's stratified queries4.623.43 (every seed)
nonuniform grid 0 / .4 / 1, K = 2, π-Flow's uniform queries3.26 ± .162.70 ± .02
K = 4 students sampled on the unseen grid 0 / .2 / .45 / .7 / 11.20 ± .061.13 ± .05 (every seed)

On the three nonuniform grids trained with the paper's grid queries, and with π-Flow's uniform queries on 0 / .4 / 1, TACD improves FID, R@1, R@3, Diversity and MM-Dist on every seed. Final segments are .52 (time-shifted) and .125 (power) long; one rule covers both. The last row uses students trained with π-Flow's uniform queries on the uniform four-step grid and sampled on the nonuniform grid without retraining. We will add these rows to the appendix.

Three more flow teachers. The same rule, unchanged, on teachers of other designs: MotionHiFlow (latent DiT, three temporal stages; bound at the end of each stage), MoGeFlow-L (flow over part-level code embeddings; bound before the snap to codes) and MARDM (we distil only the flow head inside its 18 masked-autoregressive steps; the rest stays the teacher's). FID with each model's own evaluator; "teacher" is the teacher itself sampled with that many steps; students are three-seed means after 8k updates:

modelstepsteacherno boundTACD
MotionHiFlow14 calls (default).032
6.043.042.034
32.031.89.045
MoGeFlow-L96 (default, our run).071
8.630.190.090
41.451.51.143
MARDM (flow head)adaptive (default).125
8.191.129.144
4.230.192.143

Every difference between the two student columns has the same sign on all three seeds. At every step count the TACD student beats the teacher sampled with the same number of steps, and its gain over the unbounded student shrinks as the step count grows. On MARDM's head at eight steps, where the unbounded student is already near the teacher's adaptive default, the bound costs .015.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (1/2): new experiments, and π-Flow's own query schedules

We thank the reviewers for detailed and constructive reviews. Two questions run through the reviews: whether the failure we study comes from our deterministic supervision grid (π-Flow samples query times at random, and its released configs use a 3/128 micro-window), and what a bound tied to the student's step count adds over the fixed constants used for stabilization in prior work. This comment answers the first, the next comment the second; each thread has the reviewer-specific results.

Setup of the new runs, unless stated otherwise: the paper's HY-Motion Lite student and teacher, data-free prompts, the same number of teacher calls per update, 8k updates, and the paper's evaluation (real-motion reference, 20 evaluator repetitions), at K = 8, 4 or 2 student steps, with three training seeds (42-44) per cell. "Every seed" means the paired difference (same seed, same schedule) has the same sign on all three seeds. TACD caps only the final segment's queries at its midpoint, 1 − Δt_last/2 (1 − 1/(2K) on uniform grids); every other segment is unchanged.

π-Flow's query schedules. Queries drawn uniformly within each segment (π-Flow, Alg. 1) or stratified within each segment, with and without the step-tied bound. FID ↓, mean ± SD over seeds:

queriesK = 8K = 4K = 2
uniform.765 ± .0151.168 ± .0574.77 ± .23
uniform + TACD.755 ± .032 (tie)1.080 ± .046 (every seed)3.53 ± .05 (every seed)
stratified.742 ± .0171.086 ± .0434.62 ± .23
stratified + TACD.700 ± .009 (every seed)1.003 ± .067 (every seed)3.43 ± .08 (every seed)

At two steps R@1, R@3 and MM-Dist improve with FID on every seed for both schedules. Pre-clip gradient norm (training clips at 1) over all 8k updates, ranges over seeds:

queriesupdates with b > 4K² (K = 8)P99, K = 8P99, K = 2
uniform93%249-2561,058-1,490
uniform + TACD076-88129-136
stratified100%231-2471,224-1,479
stratified + TACD073-78133-149
  1. Random queries do not avoid the near-clean weight; they hit it less regularly. The paper (Secs. 1-3) contrasts its grid with random sampling, which draws a heavily weighted time only occasionally. That holds for the extreme weights, but each update draws several queries inside the final segment, so most updates contain a query whose coefficient b = (1 − s)⁻² exceeds 4K², up to b ≈ 10⁶; we will update this contrast in Secs. 1-3 with these statistics. The bound cuts the 99th percentile of the gradient norm about 3× at eight steps and 8-12× at two.
  2. The FID damage grows as the step count drops. At eight steps the bound ties with uniform queries and gains 6% with stratified ones; at four steps it gains 7-8% on average and at two steps 26%, with the same sign on every seed for both schedules. With π-Flow's released micro-window at two steps, bounding only the final segment improves FID from 3.62 ± .16 to 3.22 ± .04, on every seed (comment 2/2).
  3. Random queries are the better schedule, and the bound composes with them. Stratified + TACD reaches .700 ± .009 at eight steps (the paper's grid: 1.04), our best eight-step result and the recipe we recommend. With these query schedules the bound is also the robust choice: it lowers FID on every seed in five of the six cells above and ties in the sixth (uniform queries, eight steps), and it makes the gradients steadier (item 1). On the paper's grid it stays in the best group at eight and at two steps, while several fixed constants that tie with it at eight steps lose 15-58% at two (comment 2/2). The deterministic grid is the controlled setting in which the terminal query is fixed and the bound can be isolated (Table 5, Fig. 6).

So the grid makes the failure extreme (2.49 vs 1.04 at eight steps); it does not create it, and under π-Flow's queries the bound's FID gain grows the fewer steps the student takes.

Also new (details in the threads): the three ports retrained at twice the budget and CD at ten times; the flow rule at two steps and on nonuniform grids; endpoint vs noise vs velocity matching on the same DDIM states; latency and quality against published efficient generators on one GPU under one protocol; a Kimodo caption check (none of the 607 evaluation captions appears among the students' training captions, verbatim or after normalisation) and results on the 99 least similar to them; training cost; π-Flow's GM policy; data-seeded rollouts; motion-level physical metrics.

AuthorsOpenReview ↗

General response (2/2): why the bound follows the step count

The inverse-square factor and its clipping are prior work; we do not claim either. Prior stabilizers bound it with fixed constants. Written as bounds on b = (1 − s)⁻²: FrameFlow and MultiFlow clip t at 0.9 (b ≤ 100), JiT uses t_eps = 0.05 (b ≤ 400), Phased DMD clamps at 10, and π-Flow's released window gives b ≤ 1,820 on our time axis. The question is what a step-tied bound adds. Same student, teacher, deterministic grid and budget; FID ↓; ± = SD over three seeds; † = seed 42 only (at K = 2 its % is against the grid rule's seed 42, 9.73); other cells are the paper's three-seed means:

Table G1-a

bound on the terminal weight bb ≤K = 8K = 2
at or below 4K² (TACD)
grid rule δ = 1/(2K) (paper default)4K²1.04 ± .039.37 ± .32
coefficient cap at 4K² (Table 5, weight capping)4K²1.019.89†
endpoint matching (Table 5)11.039.11†
fixed constants
clip at 100 (FrameFlow, MultiFlow)1001.01†10.76 ± .12 (+15%, every seed)
clip at 400 (JiT)400–12.09 ± .17 (+29%, every seed)
margin δ = 3/32 (Table 18)1141.03†9.49 ± .12 (tie)
margin δ = 1/32 (Table 18)1,0241.00†12.38 ± .22 (+32%, every seed)
cap at 4,096 (Fig. 6)4,0961.07†15.39† (+58%)
no bound≈10⁶2.4916.44†

What an eight-step sweep misses. At eight steps every fixed constant we tried lands at FID 1.00-1.07, against 1.04 ± .03 for the step-tied bound. On the same grid at two steps they separate. The margin that looks best at eight steps (δ = 1/32, 0.996) is 32% worse than the step-tied bound, JiT's clip at 400 29% and the clip at 100 15%, each on every seed and on all five metrics (R@3 .532 / .537 / .551 vs .568); the cap at 4,096 is back near the unbounded student. δ = 3/32 stays comparable, and at eight steps nothing tells it apart from δ = 1/32. So the eight-step plateau does not tell which fixed constants survive fewer steps. The failure is also silent: no update overflows, and the clip-100 student's gradient norms match ours (median 7.4 vs 7.2), so an overflow check sees nothing. Every bound at or below 4K² in Table G1-a (the grid rule, a 4K² coefficient cap, endpoint matching) stays in the best group at both step counts, with nothing tuned.

Table G1-b: π-Flow's released recipe (random queries, 3/128 micro-window), three seeds; R@1, R@3 and MM-Dist are three-seed means and improve on every seed:

K = 2FID ↓ s42 / s43 / s44mean ± SDR@1 ↑R@3 ↑MM-Dist ↓
released window3.81 / 3.50 / 3.563.62 ± .16.395.6973.77
final segment bounded by 4K²3.26 / 3.23 / 3.173.22 ± .04.405.7093.69

Why 4K². In the student's update (Sec. 3.1), step k uses the velocity (x̂₁ − x_k)/(1 − t_k), so an error in the clean-motion prediction enters that velocity with squared weight (1 − t_k)⁻², the same form as the coefficient b = (1 − s)⁻² that velocity matching gives a query at s (Sec. 3.2). On the uniform grid, every query before the final step has b at most 4 times its own step's weight (the worst case is the end of the second-to-last step). The midpoint rule gives the final step the same factor, b ≤ 4K² = 4(1 − t_{K−1})⁻², at every K, so there is nothing to tune. A fixed bound C gives the final step C/K² instead, a factor that grows 16-fold from eight steps to two (the clip at 100: 1.6 at K = 8, 25 at K = 2). This bounds the coefficient ratio, not quality directly; in Table G1-a every rule within the factor 4 stays in the best group, while beyond it results vary.

Robustness. We use the same midpoint rule, δ = Δt_last/2, at every step count, with no per-step-count tuning; on uniform grids this is δ = 1/(2K). For the flow teacher it cuts FID by 58-59% for the eight-step student in each of its three forms (Table 5), from 3.29 to 1.46 for the 1B four-step student (Table 15), and sits on the plateau of the coefficient sweep (Fig. 6, Table 18). On HY-Motion, the new runs with the paper's DX policy extend this to four and two steps, nonuniform grids and π-Flow's uniform and stratified queries: it is never worse than the same schedule without it beyond seed noise, and under these queries its FID gain grows as the step count drops. On three more flow teachers (MotionHiFlow, MoGeFlow-L and MARDM's flow head) the gain likewise grows as the step count drops. We will add Tables G1-a/b and the schedule table to the appendix, this derivation to Sec. 3.3, and FrameFlow, MultiFlow and Phased DMD to Sec. 2 next to JiT and Hong et al. (2026), which the paper already cites for the factor and for capped weights.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.