acceptodds
Under review as a conference paper at ICLR 2027

Metric-Level Universal Kinematic Attribute Control for Text-to-Motion Generation

Abstract

Text-to-motion (T2M) generation has advanced rapidly, yet current systems treat text as a soft semantic specification rather than a metric command: a model can respond to “jump high” but not to “jump 0.4 m”, and writing the number into the prompt has virtually no effect. We propose Metric Attribute Guidance (MAG), a training-free, sampling-time method that equips a frozen vanilla T2M diffusion model with a metric-level control interface over eight universal, body-intrinsic kinematic attributes, seven of which are enforced by differentiable guidance, commanded in absolute physical units alongside the caption. At each late reverse-diffusion step, MAG differentiably recovers world-space joints from the model's posterior mean, extracts soft attribute proxies, and takes a few band-limited, clipped gradient steps toward the commanded values, requiring no fine-tuning and no backbone modification. Applied unchanged to independently trained backbones, MAG raises the mean commanded-realized correlation on HumanML3D from 0.02 to 0.83 under identical commands, with no drop in text fidelity. It transfers to KIT-ML under identical hyper-parameters, reaching centimetre-level median error on the vertical extrema on both benchmarks. These results indicate that metric-level controllability is already latent in standard T2M diffusion models, and can be unlocked purely at sampling time, turning T2M from a semantic sketch into an interface that accepts commands in physical units.

Primary area:
generative models
Keywords:
text-to-motion, attribute control, training-free, gradient guidance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dnofOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Metric Attribute Guidance (MAG), a sampling-time method for controlling numerical motion attributes in a frozen text-to-motion diffusion model. MAG recovers world-space joints from the diffusion posterior mean, computes differentiable attribute proxies, and applies gradient updates with temporal DCT projection, gradient clipping, and delayed guidance. The interface exposes eight attributes, seven of which are actively guided; tempo is excluded from guidance. Experiments on independently trained MDM models using HumanML3D, KIT-ML, and an internal Basketball dataset show improved correlations between commanded and realized attributes, with centimetre-level errors for some vertical attributes. The central contribution is an interface that expresses motion requirements as measurable numerical targets. The results support useful numerical steering of selected attributes. However, they do not yet establish a broadly applicable interface that achieves accurate physical commands while preserving valid human motion.

Critical strengths
  1. The problem formulation is practically meaningful. Specifying jump height, average speed, or path length directly provides an intuitive complement to language and spatial constraints. These quantities also make control outcomes explicitly measurable.

  2. The method addresses a concrete failure mode. Gradients of extremum-based attributes can concentrate on individual frames and produce temporal spikes. The DCT projection has a clear motivation and supporting evidence: Table 15 shows that removing it increases reach Peak Jerk from 221 to 3518.

  3. The experiments include useful comparisons and diagnostics. The paper evaluates numerical prompting, attribute-conditioned training, OmniControl, and DNO using the proposed attribute objective. The appendices additionally examine attribute leakage, joint commands, retrieval under permuted commands, and skeletal deformation. These analyses expose important limitations rather than hiding them.

Critical weaknesses
  1. The attribute definitions do not adequately support the claimed universal physical semantics. Jump height and crouch depth use the median root height of the entire clip as their baseline. Appendix A explicitly acknowledges that a person who remains crouched for most of a clip and briefly stands up can receive a positive jump-height measurement and approximately zero crouch depth, even when the feet never leave the ground. This affects the meaning of the interface itself. A command to “jump 0.3 m” may be satisfied by raising the root 0.3 m relative to the clip’s median height, without performing the intended jump. Likewise, sustained crouching may not receive the expected crouch-depth measurement. Calibration on clips dominated by a neutral stance does not establish validity for these other motion categories. The evidence therefore supports control of particular trajectory descriptors more strongly than universally meaningful jump and crouch commands.

  2. Accurate attribute measurements can be achieved through invalid body deformation, and the motion-quality evidence remains incomplete. Using the same analytical function to define and evaluate a target is reasonable. The concern is that this function cannot distinguish legitimate motion from deformation. Appendix J provides a direct example: the eighth HumanML3D reach command has only 0.020 m error while the arm chain elongates by 17.9% on average. Although the authors classify these upper commands as stress conditions, Table 15 also shows a substantial quality trade-off for jump commands within the stated feasible range. MAG has a Foot Skate ratio of 0.57, compared with 0.25 for Unguided, and Peak Jerk of 5866, compared with 2578. Comparisons at matched realized jump heights are needed to distinguish effects of motion amplitude from artifacts introduced by guidance. Consequently, low attribute error and preserved text retrieval do not establish successful execution of a physical command. The central claim requires evidence that the same generated samples jointly satisfy numerical accuracy, skeletal consistency, contact plausibility, and caption semantics.

  3. High correlation does not establish accurate numerical control, and the main control evaluation has limited coverage. In Table 2, MAG’s response slopes for path length and net displacement are only 0.42 and 0.52, with median absolute errors of 1.34 m and 0.60 m, respectively. Speed error is 0.49 m/s, and facing-change error is 45.6°. These results demonstrate responsiveness to commands, but substantial response compression and numerical error remain. Facing control deteriorates further on KIT-ML, where the median error reaches 95.8°. Moreover, the main control curves largely rely on one fixed caption per attribute, with nine command levels and eight samples per level. Appendix F reports that the central speed-command range spans approximately 0.44 m/s, while the realized spread at individual levels is approximately 0.2–0.6 m/s. A high correlation across a much wider range, including extreme commands, therefore does not demonstrate reliable fine-grained control within the normal operating range. Retrieval experiments over the full test split broaden the assessment of text alignment, but do not establish command accuracy across diverse captions and action categories.

Questions for authors

Refer to Critical Weaknesses

2 comments

We thank you for the most precise diagnosis of the paper's weaknesses; each point led to a revision or a new measurement. The revised PDF is uploaded with every change highlighted in yellow, and the numbers below refer to it.

W1 — jump and crouch semantics. Agreed. With a clip-median baseline these are root-displacement descriptors. The evaluation now counts a jump only if a flight phase is detected (all four foot joints ≥ 6 cm above rest for ≥ 2 frames with the hip above standing height; on ground truth it fires on 89% of jump clips and 1% of arm-raise clips). The height is also realized in that flight: in 78% of in-range MAG samples the hip height within the flight phase is within 5 cm of the command (85% with kinematic terms). Measured from the fixed standing height instead of the clip median, jump still follows the command (ρ 0.87, median error 3.1 cm), but crouch does not (ρ 0.11): the caption already produces a sustained crouch, and the command sets the additional dip below it. Your reading is correct for crouch, and the revision states that crouch depth is a relative descriptor. Where: §3.1; Table 2 caption; Appendices A–B.

W2 — deformation, contact, matched height. Your example (eighth reach level: 0.02 m error, arm +17.9%) was real: guidance could move feature channels that are constant in the training data, such as the pelvis offset. Holding them fixed lowers the in-range torso deviation on jump from 9.7% to 1.4% and the leg deviation on reach from 18.8% to 0.8%; on reach the median error rises from 1.1 to 4.0 cm, which we report. Where: Appendix B.

We adopted your criterion as joint success on the same sample: a clip passes only if it is accurate (5 cm for heights, 0.25 m for path and displacement, 0.1 m/s, 15°), its five bone chains and its foot contact are no worse than the unguided backbone's against the fixed dataset skeleton and real-data references, and, for jumps, it shows a flight phase. In-range commands, one fixed caption per attribute (Table 6):

SpeedPathDisp.FacingJumpCrouchReach
Unguided0.030.070.120.000.120.070.12
MAG0.300.550.500.750.530.170.50
MAG + kin.0.170.380.400.820.600.620.47
DPS0.550.500.570.500.530.530.85
DNO, 100-step0.950.950.800.550.500.650.95

"+kin." adds bone-length and ground-penetration terms to the objective; they help DPS as much as MAG. Over 20 natural captions per attribute, MAG + kin. passes on 0.49–0.83 against 0.09–0.23 unguided (Table 9). With fixed real-data thresholds instead (Table 7), pass rates fall for every method and the backbone itself rarely passes; MAG's main remaining failure is foot skating on locomotion commands, which the kinematic terms do not target, and a contact term is the direct next step. Caption semantics is assessed at the set level rather than per sample: R@1 is unchanged under commands (Table 1), also with commands taken from unrelated clips (Table 17). Where: Appendix B, Tables 6–7, 9; Appendix G.

On the jump trade-off you cite (foot skate 0.57 for MAG vs. 0.25 unguided): that unguided column is numeric-prompt sampling with a different caption. With the same jump caption, unguided sampling skates 0.57 and MAG 0.59. Binned by realized jump height, MAG shows no extra degradation over the unguided backbone (skate 0.49–0.58 vs. 0.55–0.56 between 0.2 and 0.4 m); the bins hold 4 to 29 samples, so this does not exclude a contribution from the update. Where: Appendix B, Table 8.

What remains: at the two stress levels the error stays small but the motion is not physical (reach stretches the arm chain by 44% and 93%; jump sinks the body), and the reach→jump coupling (ρ 0.91) comes only from those levels. Where: Appendix E, Table 12; Appendices G, K.2; supplementary fig_operating_range.

W3 — correlation vs. accuracy, compression, single caption.

  1. The compression came mostly from evaluating the attribute objective over the padded sampling tensor rather than the clip's valid frames, which the revision corrects (original protocol; α / ρ / P50):
submittedrevised
Speed0.79 / 0.92 / 0.49 m/s0.85 / 0.93 / 0.29 m/s
Path0.42 / 0.97 / 1.34 m0.97 / 1.00 / 0.12 m
Net disp.0.52 / 0.97 / 0.60 m0.99 / 1.00 / 0.07 m
Facing0.40 / 0.63 / 45.6°— / 1.00 / 5.3°
Jump0.58 / 0.76 / 3.7 cm0.99 / 1.00 / 1.7 cm
Crouch0.49 / 0.54 / 10.2 cm1.03 / 0.99 / 1.9 cm
Reach0.89 / 0.98 / 1.8 cm0.90 / 0.99 / 2.7 cm

Where: §4.3, Table 2.

  1. The submitted levels came from whole-clip statistics (2–10 s) while clips are generated at 3 s, so the top path command (24 m in 3 s) was unattainable. We add a duration-matched protocol with levels from 3-s windows, and report per-level errors. Speed is weak on the fixed jogging caption: guidance cannot slow it below about 0.7 m/s. Where: Appendix B, Table 5; Appendix G, Table 18; Appendix K.1.

  2. Over 20 natural HumanML3D test captions per attribute (5 in-range levels × 2 seeds), in-range median errors are 0.02 m/s, 0.12 m, 0.07 m, 5.3°, and 1.3, 1.5, 2.4 cm (speed, path, disp., facing, jump, crouch, reach). Where: Appendix B, Table 9.

  3. KIT-ML facing: with joints recovered in their native units before conversion to metres (recovery integrates the root's angular velocity, so the earlier scale also rescaled the heading), it reaches ρ 1.00, median error 4.0° (95.8° submitted). Where: §4.4, Table 3.

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

The paper proposes Metric Attribute Guidance (MAG). This is a training-free method for controlling text-to-motion generation (motion control). The method defines 8 motion attributes, including speed, path length, displacement, jump height, and reach height. MAG uses differentiable attribute functions to guide a frozen diffusion model during sampling. Experiments on HumanML3D and KIT-ML demonstrate improved commanded-realized correlations compared to the vanilla text-to-motion model, while maintaining comparable text alignment.

Critical strengths
  1. The paper identifies an interesting limitation of existing text-to-motion models: natural language alone does not provide reliable numerical control over generated motion attributes.

  2. The experimental evaluation also includes a comparison between training-time attribute conditioning and inference-time guidance.

Critical weaknesses
  1. The novelty and necessity of the proposed inference-time optimization are not sufficiently justified. Recent works such as ACMDM and ProjFlow have already explored efficient motion control without inference time optimization. This raise the question of whether MAG's iterative gradient-based optimization is necessary at all. These method also shows incredible quantitative results with very little inference time significately different from the paper's Training Attribute ablation studies. I would encourage the authors to provide stronger justification for their optimization-based design and more comprehensive comparisons with recent controllable motion generation methods.

  2. The metric-level control accuracy is less convincing than the reported correlations suggest. For example, the path-length control has a correlation of 0.97 but a slope of only 0.42 and a median absolute error of 1.34 m (Table 2). Since the primary motivation is precise control in physical units, I think these results reveal limitations that are not adequately reflected by the average correlation metric.

  3. The claimed universal controllability is not fully supported by the experiments. The paper introduces eight universal kinematic attributes, but only seven are supported by differentiable guidance.

  4. The generalization results on KIT-ML are also mixed, especially for facing change and crouch depth. These results suggest that the proposed framework is less universal than claimed.

Questions for authors

Please see the weakness section and address accordingly.

1 comment

We thank you for the constructive review. The revised PDF is uploaded with every change highlighted in yellow, and the numbers below refer to it.

W1 — necessity of optimization; ACMDM and ProjFlow. ProjFlow enforces linear observations y = Ax by closed-form projection; our targets are aggregates and extrema over frames, which are not linear in the motion. ACMDM's control variant trains a ControlNet branch for spatial targets. We compare with these at the level of interface and mechanism; our empirical comparison with a spatial controller is OmniControl, which is exact where a command coincides with its (joint, frame, xyz) interface and one-sided on extrema (Table 2; Appendix I). Where: §1; §4.3.

On necessity, the evidence gives a sharper answer than our submission did. The part that carries over across solvers is the objective in physical units: given it, standard guidance (DPS, newly added with a tuned step size), noise optimization (DNO) and MAG all control the attributes. They differ in what else they preserve. Same 500 test texts (vanilla R@1 0.452):

SolverControl (mean ρ)R@1Sec/clipPeak jerk, reach
MAG0.9870.46610.0260
DPS (tuned)0.8580.3267.3296
DNO, 100-step0.9860.42342.31110
DNO, 10-step0.9780.268*6.53467

*10-step DDIM sampling of our checkpoint (trained for 1,000 steps) reaches 0.136 without any optimization, so this value reflects the sampler.

In the settings we evaluated, MAG is the only training-free solver that keeps text fidelity at this control accuracy, and the smoothest on sustained poses; DNO comes closest at four times the time, and it has lower median errors on five attributes, so we do not claim a better optimizer. Trained controllers are faster at sampling time (our attribute-conditioned backbone: 1.8 s per clip) but need a training pass per attribute set, and our attribute-conditioned backbone with guidance (AttrCond+G) lowers R@1 to 0.444 (vanilla 0.450, full test set). Where: Abstract, §1, §4.3, Tables 1–2; Appendix J, Table 20.

We also attached the objective to ACMDM itself. This is a transfer test, not a comparison with ACMDM's control variant: vertical attributes transfer to centimetres with R@1 and FID unchanged (0.489 vs. 0.483; 0.20 vs. 0.19), trajectory attributes do not (see "Additional results: ACMDM"). Where: Appendix L, Tables 22–23.

W2 — slope 0.42, P50 1.34 m. The compression came mostly from evaluating the attribute objective over the padded sampling tensor rather than the clip's valid frames. With the objective restricted to valid frames (and the method otherwise unchanged), path length is tracked with α 0.97, ρ 1.00, P50 0.12 m, and net displacement with 0.99 / 1.00 / 0.07 m. The submitted command levels also included a 24 m path in 3 s, which no clip can attain, so we add a duration-matched protocol with levels from 3-s windows. Where: §4.3, Table 2; Appendix B, Table 5.

W3 — "eight universal attributes". Agreed. The revision says "seven guided attributes; tempo is measured only", and "Universal" is removed from the title. Where: Title, Abstract, §1, §5.

W4 — KIT-ML (facing, crouch). KIT-ML joints are now recovered in their native millimetre units before conversion to metres; recovery integrates the root's angular velocity, so the earlier scale also rescaled the heading. With identical hyper-parameters (α / ρ / P50):

Attributesubmittedrevised
Facing0.00 / −0.07 / 95.8°— / 1.00 / 4.0°
Crouch0.29 / 0.38 / 20.2 cm1.01 / 1.00 / 1.6 cm
Reach0.72 / 0.81 / 4.7 cm1.00 / 1.00 / 0.5 cm
Net disp.0.45 / 0.93 / 0.71 m0.96 / 1.00 / 0.06 m
Jump0.94 / 1.00 / 2.3 cm0.98 / 1.00 / 1.5 cm
Speed0.37 / 0.98 / 0.53 m/s0.61 / 1.00 / 0.30 m/s
Path0.26 / 1.00 / 1.09 m0.36 / 1.00 / 0.91 m

KIT-ML path length remains compressed: both measurement pipelines agree on ground truth, so this is under-convergence under the shared step size, stated as a limitation. Where: §4.4, Table 3; Appendix K.1.

Reviewer_Nkd8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Metric Attribute Guidance (MAG), a training-free method for controlling metric-specific attributes in text-to-motion generation, such as jump height, movement speed, and reach height. MAG uses a frozen diffusion model and adjusts the posterior mean during sampling using differentiable attribute losses and DCT-based gradient projection. Experiments on HumanML3D and KIT-ML show that MAG improves the metrically aligned generation results.

Critical strengths
  1. The metrically-aligned T2M generation setting can be practical in some scenarios where the user wants to quantificationally control some motion attributes.

  2. The method is training-free and can be applied during inference.

  3. The experimental results show that MAG achieves more metrically-aligned generation results than the compared methods.

  4. The method is significantly faster than DNO.

Critical weaknesses
  1. Incomplete comparison with SOTA methods. For example, Table 5 only shows the comparison with relatively old methods, but recent SOTA models already achieve significantly higher results on HumanML3D. For example, Table 1 in the paper [1] show that most SOTA models achieves over 0.5 in R@1, with the best result even surpassing 0.58, while MAG only achieves 0.469 in Table 1. This raises critical concerns on the motion generation quality of MAG.

  2. The physical plausibility of the generated motions is unclear. The method directly optimizes numerical attributes without explicitly considering whether the resulting body movements are physically reasonable. For example, increasing pelvis height may lead to unnatural leg movements or incorrect foot-ground contact. This paper would benefit from fully assessing motion realism.

  3. The paper builds on existing training-free diffusion guidance methods, such as DNO, ProgMoGen, and Universal Guidance. Although the authors discuss these methods in Lines 120-122 and demonstrate a clear efficiency advantage over DNO, the proposed method itself seems relatively incremental, mainly focusing on defining a metrically measurable optimization object.

  4. It would be useful to compare MAG with standard sampling-time gradient guidance, to better understand whether the proposed optimization strategy offers benefits beyond existing guidance methods.

  5. The provided visualization examples and comparison across different models are limited in terms of both the paper and the videos in the supplementary material, making it difficult to evaluate the subjective quality of motion generation. Also, the videos do not provide the specific text prompts.

  6. The text prompts used in the paper seem to follow the fixed templates, which may limit its generalization to diverse natural language descriptions.

  7. The method is limited to predefined attributes and lacks scalability. Future work would benefit from more practical numerical control, such as walk 5 steps, jump 3 times.

[1] Li H, Lin X, Zeng L A, et al. MotionHiFlow: Text-to-motion via hierarchical flow matching. CVPR 2026.

Questions for authors

Please see the weaknesses.

1 comment

We thank you for the constructive review. The revised PDF is uploaded with every change highlighted in yellow, and the numbers below refer to it. Some numbers differ from the submission because of three implementation refinements made while addressing the reviews (the objective is evaluated on valid frames only, constant feature channels are held fixed, and KIT-ML joints are recovered in native units); they apply to every method using our objective. Where: Appendix B.

W1 — R@1 vs. SOTA. Table 1 uses a per-dataset evaluator, and MotionHiFlow's table the standard T2M evaluator, so the scales differ. MAG does not modify the network, so we test two things. (i) On the public MDM checkpoint under the standard evaluator (1,000 texts, paired seeds), MAG changes R@1/R@2/R@3 from 0.355/0.531/0.662 to 0.360/0.540/0.656 and FID from 2.68 to 3.25; our local HumanML3D copy gives GT R@1 0.458 (published 0.514), so this supports the paired comparison only. On our evaluator, R@1 is 0.467 against 0.450 for vanilla, and FID 0.063 for both. (ii) On ACMDM-S-PS22, a recent state-of-the-art-level model (unguided R@1 0.498, published 0.508), commanding the vertical attributes keeps R@1 and FID at their unguided values (0.489 vs. 0.483; 0.20 vs. 0.19); trajectory attributes do not transfer and raise FID (see "Additional results: ACMDM"). Where: §4.3, Table 1; Appendix B; Appendix L, Tables 22–23.

W2 — physical plausibility. Agreed that low attribute error does not certify a valid motion. We added a joint-success criterion on the same sample: accuracy, bone chains no more deviated from the fixed dataset skeleton than the unguided backbone's, plausible foot contact, and, for jumps, a detected flight phase. With bone-length and ground-penetration terms added to the objective, MAG passes on 0.49–0.83 of samples over 20 natural captions per attribute, against 0.09–0.23 unguided. At matched realized jump height, MAG skates and jerks about as much as the unguided backbone. With fixed real-data thresholds every method passes less often, and MAG's main remaining failure is foot skating on locomotion, which we report as the next step. Where: Appendix B, Tables 6–9.

W3/W4 — incremental; comparison with standard guidance. DNO, ProgMoGen and Universal Guidance can express a metric target once someone writes the objective; our contribution is that objective (unit-carrying, portable across datasets and skeletons) and a protocol that checks whether a command is met on the same sample, including when the number is met by deformation. Where: §1, §2.3.

You were right to ask for the comparison with standard guidance, and it turned out to be informative. We added DPS with a tuned step size, and DNO at its official 10-step setting. Same 500 test texts (vanilla R@1 0.452):

SolverControl (mean ρ)R@1Sec/clip
MAG0.9870.46610.0
DPS (tuned)0.8580.3267.3
DNO, 100-step0.9860.42342.3
DNO, 10-step0.9780.268*6.5

*10-step DDIM sampling of our checkpoint reaches 0.136 without any optimization, so this value reflects the sampler.

On the same objective all three control the attributes, so we do not claim a new optimization principle. They differ in text fidelity: MAG is the one that keeps it, which is the measured reason to use it, and the paper now states it in these terms. Where: Abstract, §1, §4.3, Tables 1–2; Appendix J, Table 20.

W5 — visualizations; videos without prompts. Figure 3 is re-rendered with SMPL meshes from the revised sweep, annotated with mean ± std over 8 samples; Figure 2 shows the full sweep for every method; the reach figures (Figures 4–5) are re-rendered the same way. New supplementary videos sweep jump, reach, speed and path length on one caption each (ladder_*), with numeric prompting above and MAG below; every clip prints the caption, the numeric prompt, the commanded and the measured value. Three new clips show Unguided, DPS, DNO, AttrCond+G, MAG and MAG + kin. side by side on the same caption and in-range command, each with that method's median-realized sample, unscreened, and its joint-success checks printed on it (compare_*). Where: Figures 2–5; supplementary videos/.

W6 — fixed text templates. We added 20 natural HumanML3D test captions per attribute, selected by a keyword rule fixed in advance (5 in-range levels × 2 seeds). In-range median errors are 0.02 m/s, 0.12 m, 0.07 m, 5.3°, and 1.3, 1.5, 2.4 cm (speed, path, disp., facing, jump, crouch, reach), and joint success with kinematic terms is 0.49–0.83 against 0.09–0.23 unguided. Where: Appendix B, Table 9.

W7 — "walk 5 steps". A count is not a differentiable function of joint positions, so it is outside the current interface; stated as a limitation and future work. Where: Appendix K.1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.