- 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
- 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.
- 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
-
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.
-
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.
- Could the authors further clarify how the observed simplicity bias can be distinguished from geometric priors learned during training?
- 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?
- 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?
- 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
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.
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.