acceptodds
Withdrawn from ICLR 2027

SCAMP: Sparse-Anchor Control Is One Small Projection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A text-to-motion generator turns a sentence into plausible human motion. Text fixes what that motion is, but not where the body must be. Authoring needs sparse anchors: chosen joints, at chosen frames, at given positions. Meeting them cur- rently costs either a conditioning branch trained for the task or hundreds of per- clip optimisation steps in the architecture’s native variables. The problem all this machinery solves is a single small one. In any generator that decodes a continu- ous state through a frozen differentiable decoder, sparse anchors define a residual whose dimension is orders of magnitude below the state’s, and meeting them is one constrained projection: the state nearest the generator’s own output whose decoded motion satisfies the anchors. Our solver, SCAMP (sparse-anchor control as a minimum-norm projection), takes that projection’s first step and continues along it: damped Gauss–Newton in the space of the anchors, through the frozen decoder alone, training-free, with one dimensionless damping constant. Applied unchanged to published generators spanning diffusion, token and latent designs, it matches or exceeds each released control method it is measured against, and closes the anchors on hosts that ship none. Holding the solver fixed turns the com- parison into an experiment on the representation, and the outcome splits along one property of the decoder: how far in time it spreads a change of state. Every gener- ator pays for the correction in foot skating; what differs is how much. A decoder that reconstructs each frame from a window of state spreads the correction over that window, and its mean contact-speed skating stays within a small multiple of the unconstrained generator’s own; an analytic recovery displaces the body itself, and skating multiplies several times over. Widening the temporal basis of the up- date lowers that cost on a frame-local decoder and raises it on one that already spreads a correction. Our generator is built to that criterion: a windowed decoder, trained through the correction, so the state it emits meets its anchors within a shal- low solve at sub-second cost. The decoder’s temporal support is a design criterion for controllable motion generation.

Primary area:
generative models
Keywords:
motion generation, motion control

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6XmsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCAMP, a training-free method for sparse-anchor control in text-to-motion generation. It formulates control as a minimum-norm correction of a generator’s continuous state subject to positional constraints on selected joints and frames. The correction is computed using a damped Gauss–Newton solver through a frozen differentiable decoder, with the linear system defined in the lower-dimensional anchor space. The same solver is evaluated on seven existing motion generators.

Critical strengths
  1. The paper offers a clean and general formulation of sparse-anchor control. Casting control as a minimum-norm correction through a frozen differentiable decoder is intuitive and potentially applicable to many motion generators without additional training.
  2. The evaluation covers a broad range of host architectures. Applying the same solver to seven diffusion-, token-, and latent-based generators is valuable and demonstrates that the proposed interface is reasonably general.
  3. The analysis of physical control cost is interesting. Going beyond anchor error to examine jerk and foot skating is appreciated. The connection between decoder temporal support and the smoothness of the resulting correction provides a useful perspective for designing controllable motion models.
Critical weaknesses
  1. Improved anchor accuracy does not consistently translate into better overall motion quality. On several retrofitted hosts, SCAMP matches or improves the control error but substantially worsens FID. For example, the reported FID is 1.05 versus 0.186 for MotionLCM and 14.2 versus 0.173 for CondMDI. This makes the headline comparison based primarily on anchor error less convincing.
  2. The “minimum-norm projection” interpretation appears stronger than what is established. Each Gauss–Newton step is a minimum-norm solution to a locally linearized constraint, but with a nonlinear decoder, the final result is not necessarily the closest feasible state to the original sample. The paper would benefit from a more careful discussion of this distinction, convergence behavior, and failure cases.
  3. The experimental setting is relatively narrow. Most experiments use HumanML3D with uniformly sampled anchors whose targets come from held-out motions. This protocol largely provides feasible, in-distribution constraints and does not sufficiently test conflicting, unreachable, irregularly placed, or out-of-distribution user specifications.
Questions for authors

1.Accuracy–quality trade-off. When lower anchor error coincides with substantially worse FID (as in MotionLCM and CondMDI), how should SCAMP's practical benefit be assessed? Please provide a systematic accuracy–quality trade-off rather than comparisons at selected solve depths only. 2.Convergence and minimum-norm projection. Since Gauss–Newton solves only a locally linearized constraint at each step, under what assumptions can the final result be interpreted as a minimum-norm projection onto the original nonlinear constraint set? Are there cases where the solver converges to a distant solution or fails to converge? 3.Robustness to challenging anchors. How does SCAMP behave with conflicting, unreachable, highly irregular, or out-of-distribution anchors that do not come from held-out ground-truth motions? These settings are important for realistic authoring applications.

Reviewer_iBPYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SCAMP, a training-free sparse-anchor control method that applies damped Gauss–Newton updates to the continuous states of pretrained motion generators. The method is evaluated across seven existing generators. The authors further investigate the influence of decoder temporal support on motion artifacts and introduce a CTMC-based generator with solver-coupled training.

Critical strengths
  1. The attempt to provide a unified sparse-control interface across different motion representations is practically relevant and is tested on seven existing generators.
Critical weaknesses
  1. The paper is excessively obscure and difficult to follow, which seriously hinders its assessment. The writing is unnecessarily convoluted, with lengthy abstract explanations that obscure the actual technical contributions. Mathematical notation is also insufficiently clarified and inconsistently contextualized, particularly the use of (\Delta) and (J) around Eqs. (3)–(4), where the evaluation points and their roles across iterations are not clearly distinguished. These presentation issues substantially undermine the clarity of the proposed formulation and method.

  2. The central theoretical claim is overstated, while the underlying optimization method offers limited novelty. Eqs. (3)–(4) minimize the norm of each local update, but do not guarantee the minimum-distance projection formulated in Eq. (1), especially for nonlinear decoders. Moreover, the core update is essentially classical damped least-squares optimization, already established in inverse kinematics. The paper does not convincingly demonstrate sufficient methodological novelty beyond applying this existing technique to motion generators.

  3. The empirical evidence does not convincingly support the effectiveness of SCAMP as a general motion-control method. Despite strong anchor accuracy, Appendix Table 7 reveals substantial degradation in generation quality: FID increases from 0.103 to 0.236 on MaskControl, 0.186 to 1.05 on MotionLCM, and 0.173 to 14.2 on CondMDI. Such degradation raises serious concerns about the practical value of the method. The evaluation emphasizes geometric accuracy without adequately establishing a favorable trade-off between control precision and motion quality.

Questions for authors
  1. How does the iterative algorithm in Eqs. (3)–(4) guarantee the minimum-distance projection formulated in Eq. (1)? Since each iteration minimizes only the local update norm and accepts steps based on residual reduction, what guarantees that the final state is nearest to the original generation? If no such guarantee exists, how should the main theoretical claim be interpreted?

  2. How do the authors reconcile the substantial FID degradation reported in Appendix Table 7 with the claimed general effectiveness of SCAMP? Can the authors provide control-accuracy versus motion-quality trade-off curves for the evaluated generators, rather than comparisons primarily based on anchor error? This is essential for determining whether the method offers a practical advantage over existing control approaches.

Reviewer_Tu1LOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper makes chosen joints of a generated motion reach chosen positions at chosen frames. It applies a few damped least-squares steps to the hidden state of a pretrained generator.

  • Input: Text, a host generator with a continuous state and a frozen differentiable decoder, and sparse anchors (six joints at eight frames in all experiments).
  • Output: A motion whose decoded anchored joints meet the targets to a chosen precision.
  • Method: SCAMP linearises the anchor residual through the frozen decoder. It takes the minimum-norm damped Gauss-Newton step Δ=−J⊤(JJ⊤+μI)−1r\Delta = -J^\top(JJ^\top + \mu I)^{-1}r and repeats it with an accept-or-reject schedule. The damping scales with the trace of JJ⊤JJ^\top. The paper uses this fixed solver to compare the physical cost of correction (jerk, foot skating) across seven hosts. It also builds a generator: a continuous-time Markov chain over VQ motion tokens, conditioned on the anchor residual and trained through the final solve.
  • Delta: The solver is the classical Levenberg-Marquardt or damped least-squares step, which the paper credits. The new parts are its use as a host-agnostic tool in generator state space, a cross-host cost study and the CTMC generator.
  • Experimental validation: On HumanML3D, the solver reaches lower anchor error than each host's released control method within one to eight steps (Table 1). Jerk and contact speed after correction separate hosts with frame-local joint recovery (21–52× jerk) from hosts with learned windowed decoders (1.1–4.6×) (Table 2). The authors' generator reaches 0.083 m in 0.50 s and 0.0026 m at FID 0.102 in 2.1 s. MaskControl reaches 0.0053 m at FID 0.116 in 9.9 s.
Critical strengths
  • The cross-host study gives a clear and useful finding about where correction costs arise. With one fixed solver, the paper isolates the representation as the variable. Hosts that recover joints frame by frame pay 21–52× jerk after correction, whereas learned windowed decoders pay 1.1–4.6× (Table 2). This matters for my rating because the finding gives a concrete design criterion for controllable motion generators.
  • The authors' generator offers a good speed and accuracy trade-off. At 0.0026 m anchor error, it reaches FID 0.102 in 2.1 s, against FID 0.116 at 0.0053 m in 9.9 s for MaskControl (Table 3). Training through the solve halves the emitted error relative to post-hoc training (Appendix H). This shows that the design criterion leads to a practical model.
Critical weaknesses

There are only two figures in the paper - and both are difficult to understand. The supplementary video is only one minute long and is not well made.

  • The writing makes the results hard to extract. The abstract and introduction use dense, metaphorical phrasing, such as "the anchors ask for little and leave most of the state free". Table 3 labels the authors' generator "SCAMP (ours)", although SCAMP elsewhere names the solver. The key quality comparison sits in Appendix Table 7.

  • As a retrofitted control method, SCAMP wins only on the metric it minimises. Table 1 and the abstract compare hosts on anchor error alone, which the solver can drive to zero by adding steps. At matched error, Appendix Table 7 shows much worse generation quality than each released method measured. FID rises from 0.103 to 0.236 on MaskControl, from 0.186 to 1.05 on MotionLCM and from 0.173 to 14.2 on CondMDI. The quality reversal appears only in the appendix.

  • The operator itself is not new. The step is damped least squares from inverse kinematics and Levenberg-Marquardt, as Section 3.2 states. ProjFlow [1] already projects generated motion onto exact spatial constraints without training, with a kinematics-aware metric that spreads the correction. The paper does not cite it.

Questions for authors

It is unlikely that I will change my rating. The paper is very badly written and very badly presented, and needs a full scale rewrite to merit acceptance.

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

Related papers

Discussion (0)

Sign in to comment.