acceptodds
Under review as a conference paper at ICLR 2027

Exploiting the Law of Prägnanz in Diffusion Models to Prevent Unauthorized 3D Structure Reconstruction

Abstract

Diffusion-based 3D reconstruction models enable accurate generation of high-fidelity 3D structures with complex geometric details based on 2D visual signals. While these models facilitate powerful design and creation applications, they also introduce significant AI safety risks by enabling adversaries to infer sensitive 3D structures, such as proprietary 3D designs or confidential physical objects. To address this threat, we present Flatland, a defense framework that protects structural information conveyed in 2D visual signals from being reconstructed by unauthorized 3D diffusion models. The key insight behind Flatland is an empirically observed simplicity bias in diffusion models: similar to human perception, their spatial reasoning behavior exhibits characteristics consistent with the psychological Law of Prägnanz, favoring simple structures over more complex alternatives. Leveraging this observation, Flatland injects simple structures into visual signals as low-amplitude perturbations, thereby disrupting unauthorized 3D reconstruction by inducing diffusion models to reconstruct the injected simple structures instead of the original sensitive ones. Experimental results demonstrate that Flatland effectively prevents 10 public diffusion models from reconstructing sensitive structures across 11 datasets. Moreover, we validate Flatland on 11 proprietary models from 9 commercial AI creation platforms, and analyze its robustness against perturbation removal and purification methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Diffusion Models, 3D Reconstruction, Protective Perturbation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_cbfUOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Flatland, a model-agnostic defense against unauthorized 3D reconstruction from 2D images, exploiting diffusion models’ preference for simple structures. It combines VLM-guided decoy synthesis, line-art-based perturbation placement, and ViT feature-space optimization. Experiments across 10 public models and 11 datasets show a 195.01% average increase in Chamfer Distance and a 44.79% decrease in F-Score. The authors also evaluate 11 proprietary models and robustness against image transformations and purification methods.

Critical strengths
  1. The paper addresses the practical challenge of protecting 2D images from unauthorized 3D reconstruction. Flatland offers an interesting approach by exploiting the observed simplicity bias of 3D diffusion models without requiring access to target models during deployment.
  2. The paper evaluates Flatland on 10 public models across 11 datasets, as well as 11 proprietary models from nine commercial platforms. The experiments also examine robustness against image transformations and perturbation removal methods, providing useful evidence of the method's effectiveness and generalizability.
Critical weaknesses
  1. Insufficient validation of the proposed simplicity-bias mechanism. The paper attributes Flatland's effectiveness to the simplicity bias of 3D diffusion models and provides supporting evidence through reconstruction experiments and neuron activation analysis. However, the observed preference for simple structures may also be influenced by learned geometric priors or differences in visual features, which are not fully disentangled. Furthermore, the neuron activation analysis in Appendix A is mainly conducted on Hunyuan3D-2.1, and the observed activation suppression provides correlational rather than causal evidence. Additional controlled experiments would strengthen the proposed explanation and clarify the contribution of simplicity bias to the method's effectiveness.

  2. Limitations in baseline comparisons and experimental fairness. The paper compares Flatland with ten existing protection methods, but most are designed for 2D image generation or manipulation rather than 3D reconstruction. More directly, Geometry Cloak [1] investigates protection against unauthorized 3D reconstruction and is cited in the submission, but is excluded from quantitative comparison due to its architecture-specific assumptions. Two other closely related studies, LFADEA [2] and Anti-Avatar [3], investigate protection against diffusion-based 3D reconstruction and unauthorized 3D head avatar generation, respectively, yet neither is cited nor discussed. Although these methods differ in their target architectures and threat models, a more detailed discussion or comparison under compatible settings would help clarify Flatland's distinct contributions. Furthermore, the paper lacks direct comparisons with simpler feature-space perturbation strategies under matched conditions, making it difficult to isolate the benefits of the proposed simplicity-bias-based design. These limitations weaken the evidence supporting Flatland's advantages over alternative protection strategies.

[1] Geometry Cloak: Preventing TGS-based 3D Reconstruction from Copyrighted Images. NeurIPS 2024.

[2] Latent Feature and Attention Dual Erasure Attack against Multi-View Diffusion Models for 3D Assets Protection. ICME 2025.

[3] Anti-Avatar: Protect Against Unauthorized 3D Head Avatar Generation via Dual-Space Divergence. AAAI 2026.

Questions for authors

Please see the Critical Weaknesses above for my main concerns and related questions.

  1. Could the authors further clarify how the observed simplicity bias can be distinguished from geometric priors learned during training?
  2. Could the authors further discuss the differences between Flatland and the closely related 3D reconstruction protection methods [1–3], and explain whether comparisons under compatible settings are feasible?
  3. The released code differs from the method described in Appendix B in several aspects, including the missing L2 regularization term in Equation (24), the similarity-based Top-K selection in Equation (21), and the SigLIP feature extraction strategy. Could the authors clarify which implementation was used for the reported experiments and provide the corresponding code, configurations and logs?
  4. Several ablation results reported in Appendix G and illustrated in Figure 8 match the baseline averages in Table 4 at the reported precision. Specifically, removing cosine distance yields 8.13%, matching AdvDM, disabling feature compression yields 4.09%, matching SDS, and removing registration yields 5.35%, matching PhotoGuard. Could the authors explain these numerical matches and provide the underlying experimental results and aggregation details for verification?

Overall, the paper is well written and visually appealing. I would be open to increasing my rating if the authors satisfactorily address the concerns raised above.

5 comments

We appreciate the reviewer for the thought-provoking suggestions. We have revised the paper accordingly, and the changes are highlighted in blue.

1.Distinguish Learned Geometric Priors and Simplicity Bias (W1, Q1)

Since 3D diffusers mainly learn complex geometries from their training datasets, the bias toward simple geometries is unlikely to originate from the learned geometric priors.

We add Fig.22(a) to compare the complex geometries in the training data and the simple geometries. The distance between the simple geometries and Trellis’s training geometries is 11.83x larger than the distance among the simple geometries (see Appendix I for other datasets). Fig.22(b) further analyzes the distribution of the geometries in visual space. Therefore, the learned priors are complex geometries and thus differ from the simple structures that the models are biased toward generating.

2.Influence of Visual Features (W1)

Thanks for the comment. To reduce the differences in visual features, we combine simple structural features with the same visual features from the complex object datasets, by instructing Gemini to carve or paint the same complex objects onto the simple objects. Fig.22(c) shows that the simplicity bias is still triggered: the distance to the complex objects is 3.19x (carved) and 3.47x (painted) higher than that to the simple objects.

We add other visual features and observe similar results (see Appendix I). These results suggest that the simplicity bias persists regardless of the visual features.

3.Controlled Activation Analysis (W1)

We add controlled experiments to strengthen the causal relationship between the simple structural signals, the suppressed activations, and the simplicity bias.

First, injecting complex structures cannot reduce the activation. We add Fig.22(d) to show that injecting complex structural signals marginally changes the activation, i.e., +2.69% in energy and +3.97% in ratio. Consequently, the signals cannot trigger the bias, and the structures change minimally (Fig.22(c)).

Second, we directly manipulate the activation to trigger the simplicity bias. As shown in Fig.22(d), injecting simple structural signals at the identified stages can reduce the activation energy by 42.54%, bringing the activation patterns close to reconstructing simple structures (gap<0.16). Consequently, the simplicity bias is triggered: the distance to the complex objects is 4.04x higher. In contrast, when the simple signals are removed at the stages, the patterns revert to reconstructing complex structures, without triggering the bias. We include detailed comparison in Appendix I.

4.Activation Analysis for More Models (W2)

We include Fig.22(e)-(g) to illustrate the activation patterns of other models: injecting simple structural signals reduces the activation energy by 8.16%, 14.02%, and 37.79% for D3S2, CraftsMan, and HY3D-2, respectively, exhibiting trends similar to Fig.10. Similar reductions are observed on Step1X-3D and Trellis, where the energy decreases by 6.48% and 22.59% (Appendix I).

5.Compare with Other Protections (W2, Q2)

We appreciate the suggestion. We enhance the comparison both experimentally and analytically.

When transferred to our 3D reconstruction protection task, Geometry Cloak, designed for a single model, and LFADEA, designed for multi-view image generation, increase the distance by only 4.73% and 3.30%, and thus fail to prevent unauthorized 3D generation (see Table 4 for details).

We further clarify the differences from these white-box methods in Table 2 and Sec.6. Geometry Cloak is architecture-specific, unlike the generic methods that we mainly compared against. Anti-Avatar is task-specific, which targets only non-diffusion facial avatar generation. LFADEA focuses on multi-view image generation, which differs from the models targeted by Flatland that generate 3D structures.

6.Extend Simpler Perturbation (W2)

We add more simple methods to ablation studies. For example, adding noise that denotes complex objects from the curated datasets decreases the quality by only 11.19% (see Table 6), while Flatland exploits the simplicity bias and achieves 17.43x higher performance. Other simple perturbations were provided in Appendix G.

7.Released Code (Q3)

Thanks for the issues. We have updated the repository.

We verified that the term is consistent with Eq.24 (opt.py:782-784) and the similarity is consistent with Eq.21(main.py:295-307). We documented the processing of remaining groups during the feature extraction in the paper. Finally, we have tested the released code and configuration by reproducing the HY3D-2.1 results in Table 1 and included the logs.

8.Numerical Issue (Q4)

We apologize for this mistake. The columns mistakenly display the SOTA baselines, whose indices collided the actual ablation baselines. We have corrected these columns, leaving others unchanged, and provide the unaggregated underlying data in Table 6.

Reviewer_cbfUOpenReview ↗

Reply to Authors' Rebuttal

Thank you for your detailed responses and code updates. I have verified that the L2 regularization and similarity-based Top-K selection are now implemented.

However, the SigLIP feature extraction still appears inconsistent with Appendix B.4. Specifically, opt.py uses the first hidden-state token as the global feature, whereas the paper describes using attention-pooled features for encoders without a CLS token. Could the authors clarify which implementation was used in the reported experiments?

Formatting Compliance. The manuscript still uses the ICLR 2026 submission header, raising concerns about compliance with the ICLR 2027 formatting requirements. Please ensure that the manuscript uses the official ICLR 2027 template and fully complies with the submission guidelines.

Thank you for your response! We have corrected the description of how our method handles ViT backbones with and without a CLS token.

Specifically, we always use the first output token of the ViT, as implemented in opt.py. For ViTs that provide a CLS token, this is the CLS token. For backbones without one, it is simply the first token. We have updated the text in Appendix B.4 accordingly (highlighted in blue).

In addition, we have updated the template. Thank you once again for all the constructive suggestions.

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

This paper introduces Flatland, which is based on and inspired by an empirically observed simplicity bias in diffusion models that is similar to the human psychology Law of Pragnanz. Flatland works by injecting simple structures into visual signals as low-amplitude perturbations such that unauthorized 3D reconstruction will be disrupted by inducing diffusion models to reconstruct the injected simple structures instead of the original sensitive ones.

Critical strengths
  1. The method this paper introduced has a clear motivation that is both empirically validated and referenced in human psychology studies, making the problem proposal interesting.

  2. The paper studies a critical problem of 3D reconstruction safety: an adversary can reconstruct proprietary 3D structures from publicly accessible high-resolution 2D images.

  3. The performance of Flatland was evaluated on 10 state-of-the-art public 3D reconstruction models with 11 datasets, making the evaluation suite complete and well-established.

  4. The threat model is clearly defined and well-established as well.

Critical weaknesses
  1. Robustness is only tested with simple transformations and noise filters, more advanced defenses against adversarial perturbations would be stronger robustness tests.

  2. Although the simplicity bias in 3D diffusion models is a good angle of motivation, the pure empirical demonstration in section 3.2 could be considered as a form of engineering effort. Is there a way to support such bias with proper theory?

  3. For the VLM used to guide simple structure synthesis, did the authors try different VLMs? How did the performance increase/decrease with each one?

Questions for authors
  1. Could the author discuss whether it would be possible for adversarial training to be applied here to any of the reconstruction models tested?

  2. Minor comment, but some of the texts in figure 3 are a bit hard to read. Could you consider increasing the font size of these texts?

  3. Minor comment: the caption of figure 4 does not provide enough information for the figure flow. Could you please expand there?

3 comments

We appreciate the reviewer for the insightful comments and questions.

1.Robustness Evaluation (W1)

Our evaluation included advanced diffusion-based noise removers from the literature. The results can be found in Figure 9 (d) and (e). Specifically, applying GrIDPure, DiffPure, and IMPRESS cannot substantially decrease protection effectiveness: the Chamfer distance degradation retains 159.66-175.86% and F-Score degradation retains 44.09-50.17%. Moreover, we tested LightShed, which trains autoencoders on protected image examples to learn and remove specific perturbation patterns. Flatland retains 159.74% Chamfer distance degradation (-18.09%) and 42.62% F-Score degradation (-4.84%).

2.Theoretical Analysis for Simplicity Bias (W2)

We appreciate the suggestion. We extend Appendix J to discuss the potential directions for theoretical analysis, based on our empirical observations.

3.Different VLMs (W3)

In Figure 8, we synthesized the structures with different VLMs (see Section 5.1). When using GPT-5 Image for synthesizing, Flatland achieved effective protection with 193.53% Chamfer distance degradation. Detailed results for using other VLMs can be found in Table 6 and Appendix G.

4.Adversarial Training for 3D Reconstruction Models (Q1)

Thanks for the question. The empirical study on adversarial training can be found in Appendix J. We applied the protective perturbations to the 100 examples for Hunyuan3D-2.1 and fine-tuned the model on these images for 10,000 steps under the default training configuration. The fine-tuned model still fails to reconstruct 3D structures, with the Chamfer distance increase remaining at 191.13%. We attributed this resilience to the fact that Flatland exploits the simplicity bias inherent in 3D diffusers, a fundamental property that may not be easily overridden by fine-tuning.

5.Minor Issues (Q2, Q3)

We are grateful for the suggestions. We have fixed them accordingly.

Thanks for the rebuttal. I remain supportive of this work.

Thank you once again for your constructive suggestions and comments.

Reviewer_p5BdOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Flatland, a model-agnostic defense against unauthorized 3D reconstruction from 2D images. Inspired by the Law of Prägnanz, the method exploits the observed tendency of 3D diffusion models to favor simple geometric structures over complex ones. Flatland uses a VLM to generate simple decoy structures (e.g., cylinders), line-art analysis to identify important regions, and ViT-based feature optimization to create protective perturbations. Experiments on 10 public models and 11 proprietary models show substantial degradation in 3D reconstruction quality, along with robustness against several preprocessing and purification methods.

Critical strengths
  1. The paper addresses an important and relatively underexplored problem of protecting 3D structural information from unauthorized reconstruction.

  2. The method does not require access to the target 3D model during perturbation generation, making it applicable to black-box models.

  3. The experiments cover a wide range of models and datasets, with extensive ablation and robustness analyses.

Critical weaknesses
  1. Insufficient justification for the Law of Prägnanz explanation. The observed simplicity bias may arise from learned model priors or visual cues rather than the proposed psychological principle. Although the authors analyze neuron activations in the denoiser, these observations do not establish a causal relationship between the simplicity bias and reconstruction failures.

  2. Limited comparison with relevant 3D protection methods. Most baselines, such as Anti-DreamBooth, Glaze, and PhotoGuard, are designed for unrelated 2D tasks rather than 3D reconstruction. The paper also cites Geometry Cloak (Song et al., 2024), a geometry-specific protection method, but excludes it from the experiments, leaving the comparison with existing 3D defenses incomplete.

  3. Limited quantitative evaluation on proprietary models. Although the paper tests 11 proprietary models, most results are qualitative, while quantitative evaluation is limited to 10 images on Meshy-4 and Hitem3D v1.0. This small-scale evaluation makes it difficult to assess the consistency of the defense across commercial models and diverse objects.

Questions for authors

Please see weaknesses.

3 comments

We appreciate the reviewer for the constructive suggestions. We have revised the submission accordingly, and the changes are highlighted in blue.

1.Distinguish Learned Model Priors from Simplicity Bias (W1)

Since 3D diffusers mainly learn complex geometries from their training data, the bias toward simple geometries is unlikely to originate from the learned model priors.

We add Figure 22(a) to compare the learned complex geometries in the training data and the simple geometries. The distance between the simple geometries and Trellis’s training geometries is 11.83x larger than the distance among the simple geometries. Detailed analysis for other datasets can be found in Appendix I. We further analyze the distribution of the geometries in visual space (Figure 22(b)). These results suggest that the learned model priors are complex geometries and thus differ from the simple structures that the models are biased toward generating.

2.Influence of Visual Cues on Simplicity Bias (W1)

We appreciate the suggestion. To reduce the differences in visual cues, we combine simple structural features with the same visual features from the complex object datasets, by instructing Gemini to carve or paint the same complex objects onto the simple objects. Figure 22(c) shows that the simplicity bias is still triggered: the distance to the complex objects is 3.19x (carved) and 3.47x (painted) higher than that to the simple objects.

We add other visual cues and observe similar results (see Appendix I). These results suggest that the simplicity bias persists regardless of the visual cues.

3.Causal Relationship Analysis (W1)

We add a series of controlled experiments to enhance the causal relationship between the suppressed activation, simplicity bias, and the reconstruction failure.

First, injecting complex structural signals cannot trigger the failure. We add Figure 22(d) to show that injecting complex structures marginally changes the activation patterns, i.e., +2.69% in energy and +3.97% in activated ratio. Consequently, the signals cannot trigger the failure, and the reconstructed structures change minimally (Figure 22(c)).

Second, we directly manipulate the activation to trigger the failure. As shown in Figure 22(d), injecting simple structural signals at the identified stages reduces the activation energy by 42.54%, bringing the activation statistics close to those of reconstructing simple structures (energy gap < 0.16). Consequently, the simplicity bias is triggered: the distance to the complex object is 4.04x higher than that to the simple object, which indicates the occurrence of reconstruction failure (see Appendix I for details).

4.Compare Other Protections (W2)

We are grateful for the suggestion. We enhance the comparison both experimentally and analytically. Specifically, we extend the comparative experiment in Table 4. Since the white-box Geometry Cloak targets only a single model, we transfer its perturbations to 3D diffusion models: it increases only 4.73% distance and thus fails to prevent unauthorized 3D generation. More comparison with other 3D protection can be found in Appendix E. In addition, we rephrase Section 6 to clarify the difference: Geometry Cloak is architecture-specific, unlike the generic methods that we mainly compare against.

5.Extend Quantitative Evaluation on Proprietary Models (W3)

Thank you for the suggestion. We add Figure 23 to extend the experiment. Specifically, we randomly sample 150 structures from the 11 datasets and submit both their original renderings and the Flatland-protected images to Tripo Studio for reconstruction. Flatland effectively prevents the reconstruction, increasing the Chamfer distance by 220.19% and decreasing the F-Score by 54.59%. We also discuss the practical challenges of evaluating proprietary models at scale, such as the substantial labor and monetary costs, in Appendix J.

Thanks for authors' response. I'll raise my score to 4.

The authors appreciate the helpful suggestions and detailed comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.