acceptodds
Withdrawn from ICLR 2027

GaitDiffusion: Diffusion-based Gait Generation Conditioned on Parkinson Severity

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating patient-specific gait patterns is essential for digital twinning and clinical assessment, data augmentation due to the scarcity of patient-specific data, and assistive robotics where gait patterns must adapt to individual pathology and therapeutic goals while preserving biomechanical realism. However, learning to generate full-body gait at clinical scale remains challenging. Here we present GaitDiffusion, a two-stage latent diffusion framework that generates full-body gait sequences conditioned on a clinical severity level scale, i.e., the Unified Parkinson's Disease (PD) Rating Scale (UPDRS). PD is a suitable testbed for pathology-conditioned generation as gait patterns degrade with clinical severity, a progression that hand-crafted healthy templates cannot represent. GaitDiffusion includes a 1D convolutional VAE with surrogate UPDRS classification and class-prototype guidance that compresses sliding-window motion into a compact latent space with embedded clinical structure, and a Diffusion Transformer (DiT) with adaptive layer normalization and classifier-free guidance, which denoises class-conditioned latents. A range-of-motion (ROM) supervision loss enforces kinematic amplitude directly on the decoded output. Trained on the multi-site CARE-PD dataset, GaitDiffusion reproduces the clinical progression of bradykinesia: mean sagittal-leg ROM decreases monotonically with severity, matching real gait with a ROM error of on average, while achieving a mean dynamic time warping of 0.585 across all four classes and an overall latent-space Fréchet inception distance of 6.28. A classifier trained on real gait features assigns the generated gait its intended UPDRS severity in 84.4% of cases, supporting use of the generated trajectories as severity-appropriate reference patterns for exoskeleton control and as synthetic training data for downstream PD gait analysis. Compared with the strongest prior pathology-conditioned generator, GaitDiffusion reduces the Average Variance Error and the Absolute Stooped Posture Mean Difference by 97.9% and 8.3%, respectively, while maintaining a competitive Absolute Arm Swing Mean Difference. Moreover, its continuous latent space enables differentiable biomechanical supervision. Overall, GaitDiffusion shows stable full-body pathology-conditioned gait generation using an ambulatory medical scale.

Primary area:
generative models
Keywords:
Diffusion models, human motion generation, Healthcare, Parkinson's disease, clinical machine learning, synthetic data augmentation, biomechanically-plausible motion synthesis

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_hL7fOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GaitDiffusion, a two-stage latent diffusion model that generates full-body Parkinsonian gait conditioned on the UPDRS gait score. The first stage is a VAE trained with an auxiliary UPDRS classification loss and a class-prototype loss. The second stage is a DiT in that latent space, trained with an additional ROM loss. On four CARE-PD cohorts, the paper reports that the per-class mean ROM of the generated gait decreases with severity as in real gait, and that the generated gait closely reproduces the real data.

Critical strengths

The problem is difficult and worth studying. Parkinsonian gait combines several changes, such as reduced stride amplitude and arm swing from bradykinesia and rigidity, stooped posture, asymmetry, and episodic freezing of gait. These differ only subtly between adjacent severity levels but vary widely across patients, and clinical data are scarce and imbalanced across levels. The paper also reports its results in detail, per class and per joint (Tables 1 and 2), which makes them easy to interpret.

Critical weaknesses

W1. Little in the method is new, and the key comparisons are not controlled.

Most components already exist: latent motion diffusion (MLD), DiT with AdaLN-Zero and CFG, and x0-prediction with kinematic losses (MDM). Benasco et al. (2026), cited in the paper, already uses the same diffusion setup for gait.

Besides the ROM loss, the new parts are the surrogate and prototype losses in the VAE, and these are not ablated (Table 4), although the conclusion says the ablation confirms their role. The VAE stage itself is not compared with diffusion directly in motion space as in MDM, and its reconstruction accuracy is not reported. The continuous-versus-discrete question is also left open, since the GAITGen numbers in Table 3 were obtained with different training data and a different evaluation protocol (Section 3.1).

Overall, it is unclear what drives the results and what the paper adds beyond existing components.

W2. The generated gait shows little diversity, so the model behaves more like a regression than a generative model.

Gait varies between patients of the same class. In the real data, the U0 to U1 difference in mean ROM (about 5.5°) is smaller than the within-class standard deviation (about 12 to 13°, Table 1), and the data augmentation that motivates the paper (Section 1) depends on reproducing this variation. The generated gait does not seem to do so. For U0–U2, its within-class ROM spread is roughly one fifth to one eighth of that of real gait (Table 1) and recall is low (Table 2), which the paper attributes to the ROM loss anchoring generation near class-mean amplitudes (Section 3.1).

A model that returns one average gait per class could also reach the low ROM error and the ROM ordering in Table 1, since these are the per-class means that the ROM loss fits directly (Eq. 6). These ROM results therefore do not separate the model from a class-conditional regression, and such samples would likely reinforce typical patterns rather than add information for augmentation.

Questions for authors

Q1 (W1). Could you add a VAE variant without the surrogate and prototype losses to the ablation, and report the reconstruction accuracy of the VAE? This would show whether these losses are what produce the ROM ordering.

Q2 (W2). Could you compare the per-class spread, not only the mean, of generated and real gait on features that the ROM loss does not use, such as walking speed, stride length, arm swing, and trunk inclination? If the generated gait matches the real within-class variation, my concern in W2 would be reduced.

Q3 (W2). Could you clarify how ROM is defined and from which part of the representation it is computed, what form the generated samples take when fed to the TRTS classifier, and what DTW is computed against? Could you also report subject-level results and confidence intervals over repeated seeds?

3 comments
AuthorsOpenReview ↗

Database and novelty

Thanks so much for the fair evaluation.

W1. We are Benasco et al. and that contribution is an Extended Abstract that we uploaded to arxiv. Thus, that paper should not invalidate the originality of this work. The current presented model is an improved version based on the ideas from that proof-of-concept. Hence, while we agree that the components are not new, the way to connect them and use them to incorporate the ROM is certainly original. Anyway, we think that the parsimony of the model and its diffusion backbone makes it very powerful for future research in gait-conditioned generation.

"Table 3 was obtained with different training data and a different evaluation protocol"

We agree as we replied to Rev 1. We never understood why GaitGen is only trained with PD-GaM and tested with the others. For instance, PD-GaM only has 2 subjects at level 3. Why PD-GaM and not the others? Is this cherry-picked? Or maybe due to compatibility to previous models. Thus, we decided to use all the data that we have and split the training, validation, and test sets by subjects. This is what would really happen in a medical application. You use the maximum amount of databases that you have.

In any case, at the time of submission, GaitGen's code was not yet released, so we reported their results instead of retraining their model. To evaluate your concern, we trained GaitDiffusion with only PD-GaM and tested it on the other databases, achieving performance similar to training with all databases. However, during this rebuttal, we discovered, by checking GaitGen code, that their reported AAMD formula in the paper was different from the one in the code. So the results were misleading in AAMD. Here, you have the results with PD-GaM-only training: AVE ~equal, AAMD 4x worse, ASMD 4x better.

What is missing to evaluate is to train GaitGen with our way of using all databases, which they did not report, but this would require unfortunately more time than the one-week rebuttal as it is not our code.

Q1 (W1). Could you add a VAE variant without the surrogate. We trained the VAE variant without the surrogate and preliminary results support the ROM ordering. But this needs further time to evaluate it properly without rush.

Q2 (W2). per-class spread. We identified that the spreads are slightly collapsed, around 0.3 less than the ground truth as expected by the reviewer. This needs careful inspection to validate our statement that we are properly generating enough different gaits for synthetic data augmentation.

Thanks so much again for the review. We will address all your concerns.

Thank you for the response and for rerunning the comparison under the GAITGen protocol. I have read it carefully.

Could you clarify whether this submission extends the cited 2026 arXiv preprint or is an independent work? The preprint is listed as an ICNR 2026 paper, so I assess this submission as an independent work and evaluate its contribution on its own.

If you are able to compute any of the requested results before the discussion period ends, namely the ablation of the two losses, the VAE reconstruction accuracy, the within-class spread on features outside the ROM loss, or the items in Q3, I would be glad to see them and will take them into account in my evaluation. Until then, I am keeping my score.

Thanks so much for your willingness to take into account the new ablations in the evaluation. However, reviewer dneg is very much in favor of rejection, so nothing will change from the outcome perspective. In any case, ff we manage to deliver all the results on time, we will provide them for the sake of completeness. The GPU is running!

Reviewer_dnegOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GaitDiffusion, a DiT-based model to generate Parkinsonian Gait conditioned on UPDRS-gait score (0–3). Stage 1 is a 1D convolutional VAE. Stage 2 is a DiT that predicts clean latents, which allows using a range-of-motion (ROM) supervision loss to directly enforce clinically relevant kinematic characteristics. The paper uses a multi-site clinical PD dataset (CARE-PD) to train and evaluate the model. It is trained on four UPDRS-annotated CARE-PD cohorts. Various metrics, including conventional human motion generation metrics and clinically relevant metrics, are used for evaluation. To measure the conditioning correctness of the model, a Random Forest classifier is trained using Train-on-Real Test-on-Synth (TRTS).

Critical strengths
  • The paper addresses a critical and significant clinical problem that has not been sufficiently explored in existing research. It extends evaluation metrics from conventional human motion literature to clinically relevant metrics. The model is also trained and evaluated on data from multiple cohorts, which is proven to be challenging because of the distribution shift between cohorts collected at different sites.

  • I like the fact that paper is open about its limitations. It reports the U3 failure, that TRTS exceeds TRTR, that FID is computed in its own latent space, and that baselines were not re-run. This is a weaker strength, but it makes the paper easier to assess.

Critical weaknesses

Weakness 1: The paper does not show that it addresses data scarcity, which is its main motivation. The paper motivates GaitDiffusion as a way to address the scarcity of clinical PD gait data. This raises a basic question: if clinical data is scarce, how can a generative model be trained at th first place? Specially how can it learn the under-represented classes well enough to produce useful synthetic samples? What made GAITGen capable of learning a generative model from scarce clinical data was its Mix and Match augmentation, which its disentangled latent space makes possible. How is this handled in GaitDiffusion? Apart from a brief mention of a class-weighted loss in the results, which is not described in the methods, the paper does not explain how it handles the under-represented class.

The results suggest the problem remains. UPDRS 3 is the weakest class on almost every metric. More generally, the generated data varies much less than the real data in all classes. Since the class where augmentation matters most is the one the model generates least well, and the synthetic data has limited diversity, the paper does not demonstrate the purpose it was designed for.

Weakness 2: Several experiments are either not comparable or missing.

  • The paper has no downstream clinical severity prediction experiment to show the usefulness of the generated samples. No experiment trains a severity estimator on real plus synthetic data and evaluates it on held-out real subjects. TRTS cannot stand in for this experiment, since it only shows that a classifier assigns the intended label to generated samples.

  • There is also no clinician rating user study on the generated samples.

  • The paper mixes all four cohorts and splits by subject; it doesn’t test on held-out sites.

  • The pooled diversity ratio of 1.227 does not support the "not collapsed" claim. Every per-class ratio is below 1, so the pooled value above 1 most likely reflects classes being pushed further apart, not within-class variety.

  • The abstract and introduction motivate "patient-specific" generation. The model only conditions on four severity classes and has no subject-level control.

  • The evaluation doesn’t support the comparative or clinical claims. The comparison in Table 3 mixes setups. Datasets, splits, sequence units (3.2 s windows versus full sequences), frame rates, and possibly metric code all differ. The baseline numbers are copied from the GAITGen paper. GAITGen uses a participant-level train/test split within PD-GaM, while GaitDiffusion uses a larger training set (four datasets, including PD-GaM), and the test samples are also different. As a result, the reported improvements cannot be interpreted.

  • Several headline numbers are measured on what the model was trained to produce (i.e., ROM error is the quantity the ROM loss optimizes). Latent FID and TRTS are computed in the VAE latent space, which was trained with a UPDRS classifier and prototype loss to separate the classes.

  • There is no evaluator independent of the generator, such as a pretrained motion encoder applied to the decoded motion same as is standard in other motion generation work.

  • Many ablation are also missing. For example, what effect do the surrogate classifier and prototype loss have, given they are contribution 1 and not ablated?

Questions for authors
  • Does adding synthetic data improve severity estimation on held-out real subjects, especially F1 for U3? Reporting results for training setups with real only, real plus synthetic, and synthetic only would be helpful.

  • Does TRTS use the generated latents directly or re-encoded decoded motion? What are TRTS and FID with an evaluator not trained jointly with your VAE? Is the 10-fold TRTR split by subject? If not, the 50%-overlapping windows from the same subject leak across folds.

  • The distance term dist(Σ\Sigma‖ck − cl‖2^2) is unbounded. What keeps the prototypes from diverging?

•The related work states that conditioning on continuous clinical severity scores has not been studied, which suggests this is a contribution of the paper. However, Section 2.3 conditions the DiT on a learned embedding of four discrete UPDRS classes, which is categorical conditioning. Could you please clarify this?

  • Eq. 5 describes a learned variance; Eq. 8 uses the fixed B_t. Which one is correct?
2 comments
AuthorsOpenReview ↗

Is GaitGen the best?

We do not want to convince you against GaitGen, and it is unfortunate that you did not learn anything from reading our paper. But here you have a diffusion alternative to GaitGen, which is easier to build, train, and improve by including other terms in the cost function.

We were in close contact with the GaitGen authors to try to be fair in comparison. But unfortunately, at the time of submission, we did not have access to their code. This is not about which model is the best is about how, together, we provide a good solution to the scarcity of data in Parkinsonian gait generation. Now that GaitGen code is released, we evaluated and we got similar numbers of the ones reported, but with their protocol of training only with PD-GaM, we got the following: AVE ~equal, AAMD 4x worse, ASMD 4x better. So fairly equal.

"The paper mixes all four cohorts and splits by subject; it doesn’t test on held-out sites." GaitGen only tests the condition PD-GaM training. We think mixing cohorts is the way to go. The on held-out should be just a double check for robustness. We still don't know why GaitGen only trains with PD-GaM.

Of course, UPDRS 3 has the worst performance; we continue working on that, and this means that the problem is not yet solved. Not by GaitGen, not by GaitDiffusion yet. We do not treat the underrepresented class differently from the others.

We wrote the limitations with transparency and sincerity to build trust, and we thank the reviewer for pointing this out.

We completely disagree with the reviewer that clinical input is the only way to subjectively evaluate these types of generators. We think that the reviewer's suggestion on: "adding synthetic data improves severity estimation on held-out real subjects" is the best way to check our statements.

In summary, at this point we do not want to convince you that our model is best, and we doubt that we will change your mind with the current numbers; thus, we will not rush our analyses and will try to address all your concerns calmly, adding all types of ablations, and resubmit.

But again, here you have a diffusion-transformer backbone that does similar work to a much more complicated architecture, by introducing ROM and classifier-free guidance. And we tested this not only on Parkinson severity but also with step-size conditioning and other simple input patterns. Thus, this shows us the potential of the architecture.

Thanks again for the review and for your precious time.

Thank you for the detailed response and for running GAITGen’s released code under its original protocol. Under this like-for-like comparison, the two methods perform comparably, so the claims of large improvements in the abstract and Table 3 should be revised in a future version. I would like to clarify why GAITGen came up in my review. It was not to argue which method is better. It was to raise a more basic question about the work. The paper’s motivation is to generate data for a field that suffers from data scarcity, in order to improve predictive models. But if clinical data is scarce, especially for severe cases, how can a data-driven generative model learn those cases well enough in the first place? GAITGen is one example of a method that explicitly addresses this, through its motion–pathology disentanglement and Mix and Match augmentation. My question was how GaitDiffusion addresses it. The response states that the under-represented class is not treated differently, and the U3 results suggest the issue remains open. To clarify one more point: I did not suggest that a clinician study is the only way to evaluate these generators; it was one of several missing experiments. I agree that the downstream experiment on held-out real subjects is the most direct test of the paper’s claims, and I take the point that training on mixed cohorts is a reasonable choice. I appreciate the authors’ openness about the limitations and their plan to address these points before resubmitting. I am keeping my score for this submission.

Reviewer_m1cFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GaitDiffusion, a latent diffusion model that generates full-body gait conditioned on Parkinson’s UPDRS-gait severity (0–3). A VAE with a UPDRS classifier and prototype loss builds the latent space, and a DiT with CFG and a range-of-motion (ROM) loss generates gait. On CARE-PD, the generated leg ROM decreases with severity, and 84.4% of samples are classified as the intended class.

Critical strengths
  1. Good Motivation: Clinical gait data is scarce and imbalanced, so severity-conditioned generation is useful for augmentation and assistive robotics.

  2. Sensible Design: The continuous latent allows a differentiable ROM loss, which token-based GAITGen cannot do. The ablation clearly shows the roles of the DiT and the ROM loss.

Critical weaknesses
  1. Unclear Comparison: Table 3 copies baseline numbers from GAITGen, where models are trained and tested on PD-GaM only, while GaitDiffusion is trained and tested on four CARE-PD cohorts. These numbers are not comparable in one table. Also, GAITGen’s table reports Diversity, which is dropped here. Please use the same training and evaluation protocol.

  2. Low Diversity:In Table 1 (Mean column), synthetic ROM std is about 7× smaller than real for U0–U2 (e.g., 1.96° vs. 13.12° for U0). In In Table 2, recall (how much of the real data is covered by generated samples) is only 0.13–0.22, far below precision (0.65–0.84 for U0–U2). The samples look real but cover only a small part of real gait. The ROM loss pulls samples toward the class mean, so the model generates near-average gait. This conflicts with the data augmentation and patient-specific claims.

Questions for authors

Pls check the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.