acceptodds
Withdrawn from ICLR 2027

ARAPDiffusion: Geometry–Distribution Feedback for Deformable 3D Shape Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deformable 3D shapes can undergo large changes in pose and morphology while still preserving plausible local geometry, yet generative models often produce samples with unnatural stretching or distortion. In latent generative models, such failures can arise from two coupled sources: the decoder determines the geometry produced at a latent code, while the generative distribution determines which latent regions are actually visited. We introduce ARAPDiffusion, a latent generative framework that establishes feedback between deformation geometry and the generative distribution. First, latent codes sampled from the current diffusion model determine where as-rigid-as-possible (ARAP) constraint is applied, concentrating geometric supervision in regions that the generator actually visits. Second, generated shapes are evaluated by their deformation relative to the observed shape family, and these scores modulate the influence of generated samples during diffusion refinement. In this way, the generative distribution determines where geometry is enforced, while deformation quality in turn reshapes what the model generates. ARAPDiffusion applies to both consistently meshed shapes and unordered point clouds, and the learned prior can also be used for conditional reconstruction. Experiments across human, animal, and bone shape families show consistent improvements in generated-shape quality, distribution matching, and reconstruction accuracy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
deformable shape generation, diffusion models, inductive biases, shape deformation model

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_NRYCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary
  • deformable 3d shape generation with ARAP constraints
  • latent codes sampled from the current diffusion model determine where ARAP constraints are applied through the generation process
  • generated shapes are scored by their deformations relative to the shape distributions from the respective training sets
  • high-level seems the motivation is whether explicit ARAP constraints / supervision help to get better deformations.
Critical strengths

Interesting exploration; ARAP regularization for deformations has been around for a long time. Then 'neural' AI methods aimed to replace these by learnt deformation networks - there's a line of work that pushes towards that; this work combines these ideas in a sense that they include the ARAP constraints into the diffusion process of generative models.

Critical weaknesses
  • The results are pretty weak: the datasets are relatively basic and it's not on par with cutting-edge methods
  • The presentation needs polish; e.g., Fig. 3/4/5 -> resolution is too low / poor quality. This makes it impossible to evaluate the results - for a top-tier paper this is not enough. Fig. 1 also not super clear what is shown.
  • I would love to see a video that shows the generation process, how the diffusion operates and have a side-by-side comparison to baselines in this regard.

Overall, I don't believe the paper is ready for publication in particular due to the presentation quality and lack of convincing results and/or their visualization.

Questions for authors

not at this point; the motivation is clear but I don't think the results are there.

Reviewer_G9GpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ARAPDiffusion aims to reduce local deformation distortion in latent generators of deformable 3D shapes. It first trains a shape autoencoder and latent diffusion model, then alternates two refinements: ARAP regularization is applied at codes sampled from the current generator, and generated samples with lower family-relative deformation scores receive greater weight in diffusion refinement. The framework uses mesh correspondences for consistent meshes and GenCorres-based deformation for implicit shapes; it also applies the learned prior to conditional reconstruction. On Human and Animal mesh and implicit settings, and Bone meshes, the authors report improvements over listed baselines on reconstruction, fitting-based shape quality, and distribution distance. Relative to NoReg, the mesh results reduce er/et/dWe_r/e_t/d_W by 19.319.3\\%/32.0\\%/17.9\% on Human and 15.615.6\\%/16.1\\%/20.4\% on Animal. Conditional reconstruction is shown qualitatively, not with a quantitative table.

Critical strengths
  • The ablations test the stated feedback dependencies rather than only removing a loss. NoDeReg and NoLaReg remove one update direction; FrozenInit and MeanMatchedUniform preserve the procedure while replacing its state-dependent signal. That design directly supports the main algorithmic claim.
  • The unconditional evaluation spans two decoder types, Human and Animal families, and Bone meshes, with reconstruction, fitting-based, distribution, and nearest-training-shape metrics. The paper also acknowledges that ete_t is a fitting-based proxy and dnd_n is a proximity check rather than a coverage metric.
Critical weaknesses
  1. The feedback loop is weakly supported. Most of the gain over prior work comes from the latent-diffusion backbone “NoReg”. The added effect of the deformation-weighted diffusion refinement is incremental.
  2. The contribution is a training-loop innovation rather than a new architecture: it combines ARAPReg-style mesh regularization, GenCorres for implicit correspondence, and quality-weighted diffusion updates. Its proposed novelty is the mutual state dependence of the updates.
  3. Related Work cites GPLD3D and reward-weighted denoising but gives each only a short description. A direct comparison of the update rules would make the novelty boundary clearer.
  4. Tables 1–2 report point estimates without seed variation or uncertainty intervals. The state-dependence ablations are reported for mesh generators, while conditional reconstruction is qualitative only. An implicit-generator feedback ablation and quantitative conditional metrics would clarify how broadly the mechanism transfers and how robust the reported gains are.
  5. The nearest neighbor step is never defined. L.252 says the deformation is measured relative to the “closest observed shape in latent space”. L261-L262 says “closest training shape to mϕ(z)m^\phi(z)”, which suggests mesh space.
Questions for authors
  1. How does the coupled update differ from GPLD3D’s quality-checker-guided diffusion objective and the reward-weighted denoising approach cited in §2? Which part of the state-dependent loop is not captured by applying either idea independently?
  2. Are the main tables single runs? If multiple runs were performed, please report the seeds and variability. If they are single runs, can repeated-run uncertainty be reported for ARAPDiffusion versus NoReg on the primary metrics?
  3. Can the authors provide the core state-dependence ablation for implicit generators? That evidence would clarify whether the feedback mechanism transfers across both decoder types. Quantitative conditional reconstruction metrics would also substantiate the paper’s conditional-use claim.
  4. Please give the exact definition of SizS_{i_z}: the space (latent_mesh), the distance, which encoder is used, and whether shapes are aligned first.
  5. Because Eq. 2 updates the encoder, does the frozen diffusion model end up sampling in a latent space that has since drifted?
  6. Please report COV/MMD/1-NNA against the test set.
  7. How were HY-LoRA and BRESA trained and sampled for unconditional generation?
Reviewer_TZTvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARAPDiffusion, a latent diffusion framework for deformable 3D shapes that alternates two refinements. First, the encoder and decoder are updated with an As-Rigid-As-Possible (ARAP) decoder regularizer evaluated at latent codes sampled from the current diffusion model. Second, the diffusion model is fine-tuned on its own samples, weighted by an ARAP deformation score relative to the nearest training shape. The claimed contribution is their bidirectional dependency. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The problem decomposition is clear: decoder errors in under-constrained latent regions versus the generator assigning probability mass there.
  2. Gains over the matched NoReg baseline are significant; e.g., on Human, ete_t drops from 10.6 to 7.21 and dWd_W from 15.1 to 12.4. 3MeanMatchedUniform is a well-designed control that removes relative sample weighting while preserving aggregate weight.
Critical weaknesses
  1. FrozenInit is confounded by latent-space drift. Eq. 2 updates the encoder and training shapes are re-encoded each alternation, so a sampler frozen at initialization draws codes in outdated coordinates. The gap may reflect encoder misalignment rather than the benefit of tracking the evolving distribution, the paper’s central claim.
  2. No variance is reported. MeanMatchedUniform trails the full model by only 0.05 mm in ete_t (though 0.8 cm in dWd_W), and the rectified-flow variant AlterDiff slightly beats the default on Human ere_r, ete_t and dWd_W.
  3. Missing baselines: N(0,I)N(0, I)-sampled ARAP regularization on the same latent diffusion backbone, and GPLD3D, cited but not evaluated. The implicit setting (Table 2) has no ablations.
  4. Evaluation shares an anchor with the training signal. rdefr_{def} uses the nearest training shape, the SMPL/SMAL fit behind ete_t is initialized from and regularized toward it, and dWd_W is computed against training samples. Coverage metrics (COV/MMD/1-NNA), test-set dWd_W, and anchor-independent validity metrics are missing.
  5. Eq. 4 is unnormalized self-training without an explicit target distribution (as the authors acknowledge) or stability analysis. Fig. 7 shows one curve per setting without error bars. 6, Conditional reconstruction is only qualitative. μ\mu (Eq. 6), the finite-difference ϵ\epsilon (App. A), latent dimension, HY-LoRA details, and code are missing. Figs. 9-11 captions wrongly praise the baselines.
Questions for authors
  1. Could FrozenInit be re-run with the encoder frozen, or with a sampler refit on current encodings without geometric weighting?
  2. Could you report mean and standard deviation over at least three seeds for key ablations?
  3. Could you add the same-backbone ARAP baseline, GPLD3D, and implicit-setting ablations?
  4. Could you provide coverage metrics and test-set dWd_W?

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

Related papers

Discussion (0)

Sign in to comment.