acceptodds
Under review as a conference paper at ICLR 2027

Plan, Don’t Pose: Long Composite Motion Generation with Text-Aligned BFM

Abstract

Text-to-motion (T2M) generation has broad applications in character animation, virtual avatars, and human-robot interaction. Existing methods typically generate pose trajectories or motion tokens directly from language, forcing a single model to handle semantic interpretation, long-horizon structure, and low-level physical realization. This coupling makes them costly and often unreliable for long, compositional, or semantically dense prompts. We propose Text2BFM, a framework that performs T2M generation directly in the latent policy space of a frozen, pretrained Behavioral Foundation Model (BFM), trained only on paired text–motion data and without video generation or heavy end-to-end motion generators. A text-aligned variational behavioral bottleneck compresses BFM policy-latent sequences into compact continuous latents that are compatible with language and preserve long-horizon behavioral structure. A lightweight conditional flow model generates in this compact space, and the result is decoded into policy latents that drive the frozen BFM in a single closed-loop rollout. Under identical clause boundaries for all methods, Text2BFM raises compositional order accuracy from 0.42 (best kinematic baseline) to 0.67 and reduces boundary discontinuity by 60%.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Text-to-motion, Behavioral Foundation Models, long-horizon motion generation

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

This paper proposes Text2BFM, which generates text-conditioned motion in the policy-latent space of a frozen Behavioral Foundation Model (BFM) instead of directly generating poses. A text-aligned variational bottleneck compresses policy-latent sequences, and a conditional flow-matching model generates them from text. The method is evaluated on HumanML3D, KIT-ML, and long compositional prompts.

Critical strengths

S1. The paper presents a clear formulation for generating motion in an executable pretrained policy-latent space rather than directly in pose space.

S2. The compositional experiments show strong improvements in stage ordering and transition smoothness over the evaluated kinematic baselines.

Critical weaknesses

Most of these weaknesses are already acknowledged in the limitations section, but most of them are not sufficiently justified despite being critical to the paper’s main claims.

W1. As acknowledged in the limitations section, the paper does not compare kinematic generators tracked by the same frozen BFM or physics-based text-conditioned controllers under a shared executor. In addition, the Meta Motivo executor was pretrained on motions overlapping with the evaluation sets. Since the official Meta Motivo repository provides reproducible training code with training-from-scratch support, retraining the BFM while excluding overlapping evaluation motions appears feasible and would provide a cleaner evaluation, but this experiment is not included.

W2. The compositional evaluation lacks sufficient directly relevant baselines. Composition-specific methods are omitted even though long compositional motion generation is one of the paper’s main claims. The paper acknowledges this limitation, but does not sufficiently justify the omission of these directly relevant baselines.

Questions for authors
  • Could the authors provide stronger comparisons using the same BFM executor and evaluate a BFM trained without motions overlapping with the evaluation set?

  • Could the authors provide qualitative videos for the main ablations to better show the effect of each component?

3 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue. In this response "the BFM" is the public Meta Motivo S-1 checkpoint unless we say otherwise.

W1, Q1. Comparison under the same BFM executor. This is done for five kinematic generators (last block of Table 3, Sec. 4.3).

After execution by the same BFMR@3 ↑FID ↓
Real motion0.7251.13
MLD0.6561.245
MotionDiffuse0.691.459
ReMoDiffuse0.6781.312
MoMask0.731.25
Our generator without the BFM0.781.96
Text2BFM0.8761.172

The table shows where text alignment is lost. When real test motions are executed by the BFM, their R-Precision drops from 0.797 to 0.725. MoMask ends at the same level after execution (0.73). The drop is therefore caused by the step in which the BFM tracks a given motion, and a better kinematic plan cannot remove it. Text2BFM has no tracking step, because it generates the policy latents directly, and it reaches 0.876.

W1, Q1. A BFM trained without overlapping motions. The reviewer is right that the Meta Motivo training code is public, and we ran this experiment (new row in Table 2, details in Appendix I). We retrained the BFM on a list without the source sequences of the HumanML3D test motions (1296 of 6973 sequences removed), in both the S and the M-1 configuration, and retrained the bottleneck and the generator on the new latents (Table 6).

BFM ExecutorR@3 ↑FID ↓MM-Dist ↓
Public S-1 checkpoint0.8761.1722.498
S configuration retrained without test sources0.8611.1692.487
Public M-1 checkpoint0.8821.0932.477
M-1 configuration retrained without test sources0.8801.1042.494

In both configurations R-Precision changes by at most 0.015, and FID and MM-Dist change by less than 0.02. R-Precision and MM-Dist stay better than those of every baseline (0.807 and 2.958 for MoMask). The new Table 12 supports this result. It evaluates all baselines on the 1420 test motions that the public BFM has not seen, and Text2BFM (M-1 checkpoint) keeps the best R-Precision there (0.830 against 0.777 for MoMask). For KIT-ML such a list cannot be built at present, because its motion ids have no published mapping to AMASS files. The revision notes this explicitly for the KIT-ML results.

W2. Composition baselines. We added FlowMDM, Motion-Agent and TEACH, and we evaluated them and Kimodo at N=4 (new Table 4, Appendix D). Text2BFM-Compose (Order Accuracy 0.671 / 0.509) is above Kimodo (0.45 / 0.17) and FlowMDM (0.53 / 0.40). The LLM-based Motion-Agent achieves higher Order Accuracy (0.91 / 0.88), also when it is executed by the same BFM (0.83 / 0.65). The abstract and Sec. 4.4 now state the result relative to the kinematic composition baselines. TEACH is trained on BABEL, so our prompts are outside its training domain (0.07 / 0.01, Appendix D).

Q2. Videos for the ablations. We added videos for the ablations to the supplementary archive (folder videos/comparison_grid, one video per prompt sequence) and describe them in Appendix J. Each video shows the same multi-clause prompt in six panels: the full model, the model without the semantic loss (no_sem), the model without temporal compression (no_compress), our generator without the BFM (vae_gen), and Motion-Agent as a kinematic output and tracked by the BFM. The prompts are out of distribution for HumanML3D, and on them Text2BFM follows the text visibly better than Motion-Agent. Without the semantic loss or without the compression, the model holds one-leg balances but does not reach the lotus pose or the standing split. Without the BFM it performs a handstand but stays upright for the breakdance and yoga clauses. Only the full model follows every clause in these sequences.

The folder videos/comparison_gravity adds videos with changed environment parameters, here Moon and Mars gravity. The generated program is executed by the BFM in the changed environment, so the resulting motion is physically consistent with that environment. The same mechanism applies to other physical parameters, such as wind and loads on the limbs. A kinematic generator alone produces the same trajectory for every environment. For comparison, the videos also show Motion-Agent tracked by the BFM under the same gravity.

Thank you for the substantial additional experiments, particularly the BFM retraining and controlled comparisons. While these address several of my original concerns, Motion-Agent still achieves substantially better stage ordering and smoother transitions under the same BFM executor. Comparisons with physics-based text-conditioned controllers also remain missing. Thus, the advantages of direct policy-latent generation for long compositional motion remain insufficiently demonstrated. I therefore maintain my original rating.

Thank you for reading the new results.

Motion-Agent in composition. On the prompts of Table 4, Motion-Agent orders stages better than Text2BFM-Compose and has smoother transitions, also under the same BFM executor. Our composition result is stated only relative to the kinematic composition methods Kimodo and FlowMDM. Two differences remain relevant. Motion-Agent does not accept clause durations, while Text2BFM-Compose follows the requested duration of every clause. The prompts of Table 4 also come from the HumanML3D distribution on which Motion-Agent is trained. On prompts outside it (videos/comparison_grid), Text2BFM performs the requested skills and Motion-Agent falls back to common motions.

Physics-based text-conditioned controllers. These methods train their own controllers for their own simulators and humanoids, so they cannot be run on our BFM without retraining. We can relate our results to their published HumanML3D numbers. CLoSD reports an R-Precision Top-3 of 0.689 and an FID of 1.798, against 0.746 for real motion in its evaluation, and MoConVQ reaches 0.614 and 3.279 in the same table. The kinematic planner of CLoSD alone reaches 0.777 and 0.283, so the loss after physical execution appears there as well. Text2BFM reaches 0.876 and 1.172, against 0.797 for real motion. The evaluation setups differ in simulator, humanoid and feature conversion, so this comparison is indirect. We will add it to the paper.

Reviewer_fQtQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes Text2BFM, a text-to-motion framework that generates continuous programs in the policy-latent space of a frozen Behavioral Foundation Model (BFM), rather than directly synthesizing pose trajectories. A text-aligned variational behavioral bottleneck compresses policy-latent sequences, and a conditional flow-matching model generates programs in this compact space. For compositional prompts, independently generated clause programs are blended and executed within one continuous rollout, preserving the simulator state between actions.

The paper evaluates text–motion alignment and distributional fidelity on HumanML3D and KIT-ML, three- and four-stage compositional generation, physical-plausibility proxies, and component ablations. Its strongest empirical results concern improved action ordering and smoother transitions under supplied clause boundaries, rather than improved distributional fidelity.

Critical strengths

Text2BFM generates representations tied to executable policy commands while delegating state-dependent physical realization to a frozen, task-agnostic controller. Executing consecutive programs within one rollout naturally carries position, heading, pose, and velocity between actions, without an intermediate kinematic plan or text-specific controller training. Although the building blocks are standard, this is a meaningful system-design contribution.

Critical weaknesses

1. The experimental comparisons are insufficient to establish the advantage over closely related prior work.

My first concern is that the experiments do not adequately position the proposed framework against methods specifically designed for compositional or long-horizon motion generation.

While the paper compares against several representative text-to-motion models, there are more closely related approaches, such as TEACH, and Motion-Agent, as well as more recent unified motion models such as LLaMo, MotionGPT3. These works investigate compositional generation, sequential motion modeling, or unified motion representations, making them particularly relevant to the paper's central motivation. Although not all operate in the same generation space or use a physics-based executor, comparisons or carefully controlled evaluations against these methods would help establish the practical advantage of the proposed interface.

In particular, the main compositional results in Table 4 compare against MoMask and Kimodo, while the four-stage evaluation includes only Text2BFM variants. This limits the conclusions that can be drawn about the method's relative advantage as compositional complexity increases.

I also find the interpretation of the general generation-quality results difficult. Although Text2BFM achieves strong R-Precision and MM-Dist, its FID is substantially worse than that of several kinematic baselines. I acknowledge the authors' explanation that the frozen BFM introduces a distribution shift and imposes a substantial FID floor under ground-truth motion reconstruction. However, this does not establish superior overall generation quality or demonstrate that the distribution shift is perceptually inconsequential.

Furthermore, the BFM has been pretrained on motion sources overlapping with the HumanML3D test set, making the comparisons in Table 2 unequal in terms of prior data exposure. The authors acknowledge this limitation and report results on BFM-unseen test motions, which is appreciated. Nevertheless, without evaluating the principal baselines on the same unseen subset, the relative advantage remains difficult to interpret.

Consequently, the current experiments do not convincingly establish how Text2BFM compares with the most relevant alternatives in terms of compositional capability, generation quality, or generalization.

2. The experiments do not isolate the claimed advantage of policy-space generation from the effects of the pretrained executor.

The authors explicitly identify the generation interface, rather than the individual architectural components, as the main contribution. However, the decisive controlled comparison—a text-aligned kinematic generator followed by the same frozen BFM—is missing.

Importantly, Table 3 already includes a suitable BFM-free generator using the same bottleneck and flow-matching framework. This model achieves better R-Precision and FID than Text2BFM. Therefore, strong semantic alignment alone does not demonstrate an advantage of policy-space generation.

The crucial question is whether the reported compositional and physical-quality improvements persist when the generator representation is varied while controlling for the executor, training data, and boundary treatment. For instance, the BFM-free generator could produce kinematic trajectories that are converted into policy latents and executed through the same frozen BFM, allowing the two generation interfaces to be compared more directly.

Similarly, the compositional improvement in Table 4 is difficult to attribute specifically to policy-space generation. Text2BFM-Compose combines latent-level blending with continuous execution by the BFM, whereas the kinematic baselines rely on pose-space stitching. The substantial reduction in transition discontinuity may therefore result from the closed-loop executor or the boundary-handling procedure, rather than from generating in policy space.

Regarding physical plausibility, Table 5 reports encouraging improvements in foot skating and contact consistency. However, these are kinematic proxies rather than direct evidence of superior physical realism. Without a shared-executor comparison or additional perceptual evaluation, it remains unclear whether the reported gains reflect the proposed representation or simply properties of the pretrained controller.

I appreciate that several of these limitations are explicitly acknowledged by the authors. Nevertheless, they directly affect the paper's central claim. Overall, although I find the direction promising and the motivation reasonable, I do not think the current empirical evidence sufficiently substantiates the paper's primary contribution.

Questions for authors

1. Comparison with compositional motion-generation methods.

Could the authors provide full model results for the principal baselines and, ideally, comparisons with composition-specific methods such as TEACH, or Motion-Agent under the same clause protocol? Such comparisons would help clarify whether the proposed approach offers a meaningful advantage over existing compositional generation methods, especially as the number of stages increases.

2. Controlled evaluation of policy-space versus kinematic-space generation.

How does Text2BFM compare with the BFM-free generator from Table 3 when the latter's predicted kinematic motions are converted into policy latents and executed by the same frozen BFM?

Could the authors compare generation qualilty, compositional order accuracy, transition discontinuity, and physical-plausibility metrics under this shared executor, with comparable boundary handling?

3. Fair comparison on BFM-unseen motions.

Could the authors evaluate Text2BFM and the principal baselines on the same BFM-unseen test subset, using identical evaluation protocols? This would help clarify the effect of prior data exposure and determine whether the reported relative performance persists on motions not encountered during BFM pretraining.

4 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue. In this response "the BFM" is the public Meta Motivo S-1 checkpoint unless we say otherwise.

W1, Q1. Composition baselines and N=4. We added FlowMDM, Motion-Agent and TEACH, evaluated them and Kimodo at N=4, and executed the kinematic baselines with a frozen BFM (new Table 4, Sec. 4.4, protocol in Appendix D). All executed rows of this table, including Text2BFM-Compose, use the larger checkpoint M-1. Each cell gives N=3 / N=4.

MethodOrder ↑Trans.×10² ↓
Stitched real clips (ceiling)0.90 / 0.853.4 / 3.3
Kimodo0.45 / 0.174.4 / 4.7
FlowMDM0.53 / 0.403.5 / 3.4
Motion-Agent0.91 / 0.882.5 / 2.4
Kimodo → BFM, stitched plan tracked0.33 / 0.113.3 / 3.4
Kimodo → BFM, latent blending0.32 / 0.114.8 / 4.7
FlowMDM → BFM, stitched plan tracked0.35 / 0.183.1 / 3.0
FlowMDM → BFM, latent blending0.35 / 0.193.9 / 3.7
Motion-Agent → BFM, latent blending0.83 / 0.653.0 / 3.0
Text2BFM-Compose0.671 / 0.5094.7 / 3.7

Text2BFM-Compose orders stages better than the kinematic composition methods Kimodo and FlowMDM, both as kinematic outputs and after execution. The LLM-based Motion-Agent achieves higher Order Accuracy, also when it is executed by the same BFM with the same latent blending. We updated the abstract, the second contribution and Sec. 4.4 so that the result is stated in this form. All baselines were re-evaluated with one revised evaluation pipeline, so the values of Kimodo differ from the submitted Table 4 (Appendix D lists the changes), and the values of Text2BFM are unchanged. TEACH is trained on BABEL, so our prompts are outside its training domain, and its results are reported in Appendix D (Order 0.07 / 0.01). MotionGPT3 and LLaMo are now discussed in Related Work, and we will add MotionGPT3 to the comparison in the final version.

Comparison outside HumanML3D. Table 4 uses prompts from the HumanML3D test set, that is, from the distribution on which Motion-Agent is trained. To see how the methods behave outside this distribution, we added videos with multi-clause prompts whose texts are not taken from HumanML3D and describe skills that are rare in it, such as a cartwheel, a handstand, floor breakdance and a standing split (folder videos/comparison_grid, described in Appendix J). On these prompts Text2BFM performs the requested skills, while Motion-Agent falls back to common HumanML3D motions. For example, it lies on the floor for "handstand" and dances upright for "breakdance". On everyday prompts the two methods are comparable.

Changed environments. The folder videos/comparison_gravity shows the same generated programs executed under Moon and Mars gravity. The generator is not changed. The BFM executes the program in the changed environment and adapts the motion in closed loop, so the result is physically consistent with that environment. The same mechanism applies to other physical parameters, such as wind and loads on the limbs. A kinematic generator alone produces one trajectory for every environment. For comparison, the videos also show Motion-Agent tracked by the BFM under the same gravity.

W1. FID and overall generation quality. We agree with this point. Our claim concerns text alignment under physical execution and not overall generation quality, and the abstract and the introduction now say this explicitly. The last block of Table 3 shows that every method executed by this BFM has an FID of at least 1.13, and this includes real motion. A user study did not fit into the discussion period, and we will add one to the final version.

W1, Q3. Baselines on the same unseen subset. This is done in the new Table 12 (Appendix I). On the 1420 test motions that the public BFM has not seen, Text2BFM (M-1 checkpoint) keeps the best R-Precision (0.830 against 0.777 for MoMask) and MM-Dist (2.645 against 3.02), and its FID stays higher. We also retrained the BFM without the sources of the HumanML3D test motions (new row in Table 2). On the full test set R-Precision changes from 0.876 to 0.861 in the S configuration and from 0.882 to 0.880 in the M-1 configuration. FID and MM-Dist change by less than 0.02, and R-Precision and MM-Dist stay better than those of every baseline (new rows in Table 2 and Table 6).

W2, Q2. Control with the same BFM executor. We ran the control that the reviewer proposes (last block of Table 3, Sec. 4.3). The kinematic motions of our BFM-free generator are converted to policy latents and executed by the same frozen BFM. This generator uses the same bottleneck, the same flow model and the same data as Text2BFM, so after execution the two pipelines differ only in the space in which generation happens.

PipelineR@3 ↑FID ↓
Our BFM-free generator, kinematic output0.9120.234
Our BFM-free generator, executed by the BFM0.781.96
Text2BFM, policy-space generation with the same BFM0.8761.172
Real motion, executed by the BFM0.7251.13

The reviewer is right that before execution the BFM-free generator has better R-Precision and FID. After execution by the BFM it loses this lead. Its R-Precision falls from 0.912 to 0.78 and its FID rises to 1.96, while Text2BFM reaches 0.876 and 1.172. Four published kinematic generators show the same pattern and reach 0.656 to 0.73 after execution. The advantage of generating in the policy space therefore appears when the motion must be physically executed.

The table also shows what the BFM costs, so we state what it adds. First, the motion is produced by a physics-based policy and has fewer contact artifacts. With the same generator, foot skating drops from 0.113 to 0.034 m/s and contact consistency rises from 68.2% to 94.7% when the BFM is used (Table 5). Second, the BFM helps on prompts outside the training distribution. In the out-of-distribution videos the full model follows every clause, while the BFM-free generator misses some of them, for example the breakdance and yoga clauses (folder videos/comparison_grid, Appendix J). This second observation is qualitative.

W2. Source of the gain in Table 4. The reviewer asks whether the gain comes from the closed-loop executor, from the boundary handling, or from generating in the policy space. The lower block of the new Table 4 keeps the executor fixed and varies the other factors, so we can answer for each factor separately (values at N=3).

  • Executor. Execution by the BFM does not produce the ordering gain. It lowers the Order Accuracy of the kinematic methods (Kimodo from 0.45 to 0.33, FlowMDM from 0.53 to 0.35). It makes their transitions slightly smoother (4.4 to 3.3 and 3.5 to 3.1).
  • Boundary handling. With the same executor, latent blending and pose stitching give the same Order Accuracy (0.32 and 0.33 for Kimodo, 0.35 and 0.35 for FlowMDM). Latent blending gives less smooth transitions than tracking a stitched plan (3.7 to 4.8 against 3.0 to 3.4).
  • Generator. With the same executor and the same latent blending, Text2BFM-Compose reaches 0.671, Kimodo 0.32, FlowMDM 0.35 and Motion-Agent 0.83. The ordering result is therefore determined by the clause-level generator.

For ordering, the gain of Text2BFM-Compose over Kimodo and FlowMDM is not explained by the executor or by the boundary handling. We do not attribute it to the generation space alone, because these methods also differ in the generator itself. The controlled comparison with one generator in both spaces is available for single prompts (W2, Q2 above) and not yet for composition. For transitions, the reviewer is right. The low discontinuity of Text2BFM-Compose does not come from the policy space, since its value (4.7) lies in the range of latent blending, and the revision does not claim smoother transitions.

W2. Physical proxies in Table 5. We agree that the physical improvements come mainly from the BFM executor, and the new Table 11 (Appendix E) shows this directly. After execution by the same BFM (M-1), MoMask skating drops from 0.132 to 0.088 m/s and its contact consistency rises from 63.6% to 87.6%. The remaining difference is smaller. Text2BFM has the lowest contact slide in the simulator (0.017 m/s against 0.023 to 0.034), while executed Kimodo has slightly lower skating on the proxy (0.062 against 0.067).

I thank the authors for the extensive work! The substantial additional experiments and clarification enhanced paper presentation and soundness. The shared executor and BFM retraining results address several of my major experimental concern. And those provide valuable evidence for the benefit of policy-space generation.

However, the long-compositional claim is still a bit weak. I appreciate the experiments against Motion-Agent but they showed better results empirically. The new experiments also strengthen the physcial evaluation, however, as stated by the authors, those gains mainly comes from the BFMs. I still find the manuscript below my standard for acceptance. But this work also leads to a discussion that I find interesting: shall we operate on motion or policy space?

Overall I will retain my score at 2, while raise soundness from 1 to 2.

Thank you for the reading of the new results and for raising the soundness score. We agree with both remaining points and would like to say how we address them.

Long composition. Motion-Agent is better on this benchmark. The revised abstract and Sec. 4.4 already state this, so that composition is presented as a study and not as the main claim.

Physical gains come from the BFM. We agree, and Sec. 4.5 now says so. The contribution of policy-space generation is different. A kinematic plan that is tracked by the same BFM obtains the same physical benefits but loses text alignment in the tracking step (R-Precision 0.66 to 0.78 for five generators, and 0.725 for real motion). Generating the policy latents directly keeps the physical benefits without this loss (0.876).

Motion space or policy space? We think the experiments give a first answer to this question, and we will organize the discussion around it.

  • If the output is a kinematic motion, motion space is the better choice. Our own generator without the BFM has better R-Precision and FID than Text2BFM (0.912 and 0.234).
  • If the motion must be physically executed, policy space is the better choice for single prompts. With the same bottleneck, generator and data, R-Precision is 0.876 against 0.78 and FID is 1.172 against 1.96, and the result holds with a BFM retrained without the test-motion sources.
  • For composition, the quality of the clause-level generator dominates the choice of space.
  • The cost of policy space is tied to the executor. Its coverage sets a lower bound on FID (1.13 for real motion).
Reviewer_1h1tOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed Text2BFM, a framework that incorporates Behavioural Foundation Models (BFMs) into text-to-motion (T2M) generation tasks. BFMs learn a structured space of reusable behaviours; they are commonly used in embodied AI and can serve as an interface, preventing T2M models from generating joint-space motions directly. The paper presents this end-to-end workflow, introduces a variational behavioural bottleneck to align text with behaviour embeddings, and evaluates its effectiveness on the HumanML3D and KIT-ML benchmarks.

Critical strengths

The paper’s critical strength is the formulation of the novel Text2Motion Generation Framework. Text2BFM operates directly in policy-latent space and hands over the execution to pre-trained BFMs. This creates a meaningful interface and separates high-level semantic planning from low-level physical realization. The framework is especially effective for action sequences. Given a compositional text sentence, Text2BFM-Compose will treat it as multiple ‘clauses’ and execute in a single closed-loop BFM rollout, which can ensure semantic order accuracy and reduce the motion boundary discontinuity. The evaluations also provide strong evidence for the performance. (Table 4). The mathematical analysis also contributes to the soundness of Text2BFM’s motivation and design.

Critical weaknesses
  1. The unequal data exposure of nearly 69% of HumanML3D test motions in BFM (Meta Motivo) is discussed in the main text (Section 4.3) and listed as limitations. The authors try to evaluate Text2BFM on the remaining subset, but it cannot convince Text2BFM’s advantage on other baselines in this setting. A clean comparison is needed with no test data leakage; it is a crucial problem, not future work.
  2. The paper evaluated the performance removing the BFM, results show the FID/MM-Dist/R-Prec all improving, authors argued that the phenomenon explains the text-alignment is mainly contributed by the bottleneck and generator, the BFM costs distributional fidelity but brings physical plausibility and fewer artefacts. However, the paper only designed kinematic proxies as physical metrics, but lacked a true user study for motion quality to verify that the BFM-induced FID shift is perceptually harmless. Also, this problem should not be listed as a limitation; it is a key proof absence.
  3. The reported FID gaps shown in all BFM-related pipelines are explained as a distribution shift introduced by BFM itself. The authors then translated the ground-truth motions via BFM and got an FID of 1.13, which the authors claimed as the executor floor. It is reasonable but lacks strong evidence. An additional comparison between 1) current Text2BFM and 2) passing baselines’ generated motions through frozen BFM could be helpful to clarify the FID problem.
Questions for authors
  1. Please conduct an equal comparison and replace the current Table 1. Authors may follow the exact test split of BFM and retrain baselines with complete release codes like T2M, MDM, MoMask, etc.
  2. Please supplement the user study on the preferences between Text2BFM and other baselines from real humans.
  3. Please supplement the comparison of FIDs between Text2BFM and other baselines, but through BFM translation.
2 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue. In this response "the BFM" is the public Meta Motivo S-1 checkpoint unless we say otherwise.

W1, Q1. Test data seen by the BFM. We agree that the paper needs a comparison in which the BFM has not seen the test motions. The revision adds two such comparisons (Sec. 4.3).

First, we retrained the BFM without the source sequences of the HumanML3D test motions (1296 of 6973 sequences removed), in both the S and the M-1 configuration, and retrained the VAE bottleneck and the generator on its latents. The new rows are in Table 2 (called Table 1 in the review) and Table 6 and give an equal comparison with the baselines on the full test set.

BFM ExecutorR@3 ↑FID ↓MM-Dist ↓
Public S-1 checkpoint0.8761.1722.498
S configuration retrained without test sources0.8611.1692.487
Public M-1 checkpoint0.8821.0932.477
M-1 configuration retrained without test sources0.8801.1042.494

In both configurations R-Precision changes by at most 0.015, and FID and MM-Dist change by less than 0.02. R-Precision and MM-Dist stay better than those of every baseline (0.807 and 2.958 for MoMask). MultiModality is also unchanged (1.232 against 1.233 for S, and 3.365 against 3.371 for M-1).

Second, we evaluated the baselines of Table 2 that have public checkpoints, and Text2BFM with the M-1 checkpoint, on the 1420 test motions that the public BFM has not seen, with one pipeline for all methods (new Table 12, Appendix I). Text2BFM keeps the best R-Precision (0.830 against 0.777 for MoMask) and the best MM-Dist (2.645 against 3.02). Its FID stays higher (1.306 against 0.174 for MoMask). The baselines were trained on the training split only, so they can be compared on this subset without retraining.

For KIT-ML the same check is not available at present. Its motion ids have no published mapping to the AMASS files that define the Meta Motivo training set, so the overlap with the BFM training data cannot be measured. The revision notes this in Sec. 4.3 and Appendix I.

W2, Q2. User study and physical metrics. A user study did not fit into the discussion period, and we will add one to the final version. Until then the revision lists it as a limitation (Sec. 5) and makes no claim about the perceptual effect of the FID shift.

We agree that kinematic proxies alone do not prove physical quality. The new Table 11 (Appendix E) executes kinematic baselines with the same BFM (the larger checkpoint M-1) and adds quantities measured in the simulator. Text2BFM has the lowest slip of foot contacts (0.017 m/s against 0.023 to 0.034 for executed baselines) and the lowest share of sliding contacts (2.4% against 4.1% to 6.3%). Its fall rate in the simulator is 0.6%, equal to the best executed baseline.

W3, Q3. FID of the baselines after the BFM. We ran the suggested comparison for five generators (last block of Table 3, Sec. 4.3).

After execution by the same BFMR@3 ↑FID ↓
Real motion0.7251.13
MLD0.6561.245
MotionDiffuse0.691.459
ReMoDiffuse0.6781.312
MoMask0.731.25
Our generator without the BFM0.781.96
Text2BFM0.8761.172

Every executed kinematic generator has an FID above 1.13, the value of executed real motion. Their R-Precision after execution is 0.656 to 0.78, against 0.876 for Text2BFM. A conversion control in Appendix I shows that the conversion of kinematic outputs does not penalize the baselines.

Changes to the claims. We refined the abstract and the introduction. The main claim is that, under the same BFM executor, generating in the policy space preserves text alignment better than a kinematic plan followed by tracking. On HumanML3D, Text2BFM reaches an R-Precision of 0.876, against 0.66 to 0.78 for five kinematic generators executed by the same BFM and 0.725 for executed real motion. Its FID (1.172) is also the lowest among the executed methods and the closest to that of executed real motion (1.13). The result holds when the BFM is retrained without the test-motion sources.

Thanks for the authors' reply and additional work. Especially the extensive experiments on non-leakage BFM retraining and the comparison under the same BFM execution addressed my critical concerns. Also, the change in the claim provides a clearer scope of the work. For the user study, due to the time limitation, it is understandable to provide till the final version. Overall, I will raise my rating to 3

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.