acceptodds
Under review as a conference paper at ICLR 2027

AnchorGen: Anchored Optimization for Customizable Generative 3D Design

Abstract

Engineering design often starts from a 2D sketch that fixes style and proportions, yet the subsequent 3D shape optimization relies on learned generative priors to keep the geometry valid. However, these priors are agnostic to the sketch: while they admit a valid design by correcting a drifted proposal back to its training distribution, they often correct it towards the high-density region, ignoring the specified design. We introduce AnchorGen, a rectified-flow framework trained unconditionally on the concatenated shape and sketch latents of paired data. The learned manifold represents the joint distribution of shape-sketch pairs, so constraining the sketch component restricts the iterate to the sub-manifold of shapes consistent with a target style. Since training employs no conditioning signal, the constraint is imposed at inference: gradient descent optimizes the shape latent to minimize a differentiable drag surrogate, while constraining the sketch latent to remain close to the target sketch via a token-wise cosine penalty. A single model thereby supports design-preserving optimization, dimensionally explicit design edits, and sketch-only synthesis.

Primary area:
generative models
Keywords:
Industrial Design, 3D Generation, 3. Anchored Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2u5rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents AnchorGen, a framework for physics-driven 3D shape optimization that incorporates 2D sketches to preserve and control design intent. The method first encodes 3D shapes and their corresponding sketches into separate latent representations, which are then concatenated to train an unconditional joint flow matching model. During inference, the framework optimizes an intermediate state of the learned flow by combining a differentiable aerodynamic drag objective with a sketch anchoring loss based on token-wise cosine similarity. This allows the shape and sketch representations to evolve jointly while encouraging the generated geometry to remain consistent with a target sketch. The framework supports three applications: optimizing an existing shape while preserving its design, modifying specific geometric features through edited sketches, and generating 3D shapes from sketches without an initial model. The authors evaluate AnchorGen on ShapeNet and DrivAerNet++, comparing it with prior-based optimization methods and sketch-conditioned generative approaches. The experiments examine aerodynamic drag reduction, geometric similarity, dimensional control, and shape preservation, along with additional analyses of sketch guidance strength, model capacity, and generation from sparse sketches.

Critical strengths

This work introduces 2D sketches into physics-driven 3D shape optimization, allowing users to express their design intentions through intuitive and easily editable sketches. By incorporating sketch-based constraints into the optimization process, the method improves the controllability and flexibility of 3D shape design.

Critical weaknesses
  1. Limited Novelty. Previous studies have already incorporated physics-based optimization into 3D generation. This work mainly extends existing approaches by jointly modeling 2D sketches and 3D shapes while optimizing both sketch consistency and physical performance. Therefore, the methodological novelty appears relatively limited.

  2. Limited Generalization. Although the paper presents AnchorGen as a general framework for generative 3D design, all experiments focus on automotive shapes and aerodynamic drag optimization. The authors should consider evaluating additional shape categories and physical objectives or narrowing the scope of their claims to better align with the experimental evidence.

Questions for authors
  1. What is the specific advantage of jointly modeling 2D sketches and 3D shapes compared to simply adding a sketch consistency loss to an existing physics-guided shape optimization framework? Could the authors provide an ablation to better isolate the contribution of joint modeling?

  2. The paper presents AnchorGen as a general framework for generative 3D design, but the experiments focus exclusively on automotive shapes and drag optimization. Is there any evidence that the method generalizes to other shape categories or physical objectives? If not, could the authors clarify the scope and limitations of the proposed framework?

4 comments
AuthorsOpenReview ↗

Author Responses (1)

W1 / Q1. Joint Modeling vs. Sketch Consistency Loss.

What is the specific advantage of jointly modeling 2D sketches and 3D shapes compared to simply adding a sketch consistency loss to an existing physics-guided shape optimization framework?...

A: Our goal is to preserve the design style specified by a sketch during physics-driven optimization. Existing methods either only keep the shape on the manifold of valid shapes and ignore its style, as FCSO does, or require the user to explicitly specify which existing features to preserve, as GANO [1] does through null-space projection. Neither keeps the overall style that a sketch expresses.

As the reviewer suggests, the simplest way to achieve our stated goal would be to incorporate a sketch consistency loss into an existing physics-guided shape optimization framework, which is similar in spirit to Sketch2Mesh [2]. Unfortunately, this is not as easy or effective as it sounds:

  • No robust projector. Hard edge projections of the shape are binary and give no gradient, and blurring them to obtain one changes what is compared and loses shape details. In Appendix I, we tried this and both blurred projection variants yield much less sharp-looking results than those of AnchorGen. The metrics in the table below confirm this (Proj-Ext: exterior contour; Proj-All: all edges; same setting as Table 2).
MethodF@1% ↑[email protected]% ↑CD-L1 ↓Vol-Ratio ↑Normal-C ↑HD95 ↓Avg. Dim-Err ↓
Proj-Ext68.5247.551.86993.3488.502.9572.67
Proj-All69.4247.511.80792.8888.622.8392.63
AnchorGen (Ours)79.6855.601.42692.6291.012.3922.67
  • High cost. The projection must be rendered from the decoded shape at every iteration, which nearly doubles the time per step for the edge projection (1.13 s vs. 0.60 s) and makes a learned line-art model such as Qwen impractical (about 80 s and 70 GB of GPU memory per step).
  • Sparse and style-sensitive signal. A pixel-level loss such as MSE between line drawings gives a sparse signal, since the lines cover only a small fraction of the image, and is sensitive to the drawing style, so the same shape drawn in another style yields a large loss. In our tests with sketches of different styles, such a loss easily made the optimization collapse.

Our design avoids the first two problems and mitigates the third: the anchor is a cosine distance in the sketch latent space, so no projector or rendering is needed, and since the correspondence between shape and sketch is learned from data, comparing latent tokens rather than pixels is less sensitive to the drawing style.

[1] Guoze, Sun, et al. "Geometry-Aware Neural Optimizer for Shape Optimization and Inversion." Forty-third International Conference on Machine Learning.

[2] Guillard, Benoit, et al. "Sketch2mesh: Reconstructing and editing 3d shapes from sketches." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021.

AuthorsOpenReview ↗

Author Responses (2)

W2 / Q2. Generalization to Other Shape Categories and Physical Objectives.

The paper presents AnchorGen as a general framework for generative 3D design, but the experiments focus exclusively on automotive shapes and drag optimization. Is there any evidence that the method generalizes to other shape categories or physical objectives?...

A: The reviewer is right and we addressed this concern by adding a chair volume-reduction experiment (Sec. 4.4). In the spirit of those in FCSO, we reduce the volume of each chair by 15%, which would make it lighter, save material, and lower its production cost in real industrial settings.

  • Truly different. Unlike drag, which is predicted by a learned surrogate, volume is computed directly from the shape. Yet minimizing it can easily destroy realism: thin legs break, holes appear in the backrest, or the chair turns into a generic boxy one. This makes it a direct test of whether the optimization preserves the original design, complementing the experiments on cars.

  • Setup. A single model is trained on synthetic sketches and anchored at test time to real hand-drawn sketches of 30 held-out chairs from three views. AnchorGen again consistently stays closer to the original chair than all baselines on most metrics as reported in the table below.

Fig. 5 and Fig. 15 of the revised paper provide qualitative results showing that it preserves the initial design, whereas the baselines break thin legs, fill the backrest slots, or make the chair boxy. The table below reports the similarity of each optimized chair to its own original shape at a 15% volume reduction, separately for each sketch view (all metrics are scaled by ×100). More details are given in Appendix H.

MethodF@1% ↑[email protected]% ↑CD-L1 ↓Vol-Ratio ↑Normal-C ↑HD95 ↓Avg. Dim-Err ↓
GD (2025)94.8270.250.9084.1396.951.051.83
FMG (2022)57.3934.713.2085.4682.095.364.03
D-Flow (2024)59.3937.043.7880.9582.067.595.71
ICTM (2024)54.0532.253.1076.5280.175.816.61
SGO (2026)95.6273.580.8583.9497.580.921.62
FCSO (2026)92.6671.450.9783.8496.511.321.98
AnchorGen (0°)97.9283.220.7783.8598.130.901.05
AnchorGen (45°)97.3882.860.8383.9897.851.021.11
AnchorGen (90°)98.0382.440.7983.9598.030.951.11

Thank you for the responses and additional experiments. The new comparisons help explain why directly using sketch projections is less effective and clarify the motivation behind the joint modeling approach. The chair experiment also helps address my concern about the method's applicability beyond automotive design. While further tests on other categories would still be useful, the authors have addressed my main concerns, and I am satisfied with the revisions.

AuthorsOpenReview ↗

Thank You for Your Feedback

Dear Reviewer 2u5r,

Thank you for your positive feedback and for acknowledging our revisions. We truly appreciate your time and constructive comments.

Best regards,
AnchorGen Authors

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

This paper proposes AnchorGen to address the loss of design intent from an initial 2D sketch when optimizing physical performance, such as aerodynamic drag, in 3D engineering design. Existing optimization methods based on generative priors can correct geometry that deviates from the shape manifold, but tend to converge toward high-density regions of the data distribution rather than preserve the characteristics of a specific design. To address this limitation, AnchorGen encodes 3D shapes and 2D sketches into their respective latent spaces and learns their joint distribution using unconditional rectified flow matching. At inference time, it optimizes physical performance through a differentiable drag surrogate while preserving the target design using a sketch anchoring loss based on token-wise cosine similarity. Experimental results show improved geometric similarity and dimensional controllability over existing prior-based and conditioning-based methods, demonstrating the ability to achieve both aerodynamic performance and design fidelity.

Critical strengths

S1. By learning the joint distribution of shapes and sketches through unconditional flow matching and applying anchoring at inference time, the proposed method can flexibly combine different design constraints and physical objectives without retraining.

S2. The authors compare the proposed method with various prior-based and conditioning-based baselines on ShapeNet and DrivAerNet++. Evaluations of geometric similarity, dimensional accuracy, together with ablation studies, provide evidence of its effectiveness across several aspects.

Critical weaknesses

W1. Most experiments focus on synthetic sketches automatically generated from 3D geometry and on drag optimization for cars. Generalization to diverse freehand sketches drawn by actual users, other shape categories, and other physical objectives is therefore insufficiently validated. This is a significant limitation and my primary concern regarding this paper. The empirical evidence is very limited to substantiate the claims made in this paper. I expect the authors to address this concern thoroughly in their strong rebuttal.

W2. Sketch anchoring relies on token-wise cosine similarity, so similarity in the sketch latent space does not guarantee preservation of specific design features in the resulting 3D geometry. Moreover, the anchoring strength introduces a trade-off between design fidelity and physical optimization. Additional analysis of these limitations is needed.

Questions for authors

Please see the weaknesses section.

6 comments
AuthorsOpenReview ↗

Author Responses (1)

W1. Generalization: freehand sketches, other shape categories, other physical objectives.

Most experiments focus on synthetic sketches automatically generated from 3D geometry and on drag optimization for cars. Generalization to diverse freehand sketches drawn by actual users, other shape categories, and other physical objectives is therefore insufficiently validated...

A: To expand our method to other shape categories and physical objectives, we have added a chair volume-reduction experiment (Sec. 4.4), which provides an additional test case for both the shape category and the physical objective. The goal is to reduce the volume of each chair by 15%, with the motivation in mind to make it lighter, saving material, and lowering hence the production cost. Unlike drag, which is predicted by a learned surrogate, volume is computed directly from the shape. Yet minimizing this objective can easily break the shape: thin legs break, holes appear in the backrest, or the chair turns into a generic boxy one. This makes it a direct test of whether the optimization preserves the original design.

This chair dataset further allows us to test generalization to real sketches. We follow the setting of Sketch2Mesh [1]: the model is trained only on synthetic sketches, 32 per chair, and tested on real hand-drawn sketches. We report results on 30 held-out chairs, each anchored to real sketches from three different views (0°, 45°, and 90°), using one common model. The table below, added as Table 3 in Sec. 4.4 of the revised paper, reports the similarity of each optimized chair to its own original shape at a 15% volume reduction, separately for each sketch view (all metrics are scaled by ×100).

MethodF@1% ↑[email protected]% ↑CD-L1 ↓Vol-Ratio ↑Normal-C ↑HD95 ↓Avg. Dim-Err ↓
GD (2025)94.8270.250.9084.1396.951.051.83
FMG (2022)57.3934.713.2085.4682.095.364.03
D-Flow (2024)59.3937.043.7880.9582.067.595.71
ICTM (2024)54.0532.253.1076.5280.175.816.61
SGO (2026)95.6273.580.8583.9497.580.921.62
FCSO (2026)92.6671.450.9783.8496.511.321.98
AnchorGen (0°)97.9283.220.7783.8598.130.901.05
AnchorGen (45°)97.3882.860.8383.9897.851.021.11
AnchorGen (90°)98.0382.440.7983.9598.030.951.11

For all three views, AnchorGen stays closer to the original chair than all baselines on most metrics. More experimental details and visualizations are given in Appendix H (Anchored Optimization with Real Sketches). In addition, the experiment in Fig. 6 of the original submission already shows, to some extent, robustness to different sketch styles: when generating from scratch, AnchorGen adapts to simple line drawings that it never saw during training (Fig. 6(b), zero-shot).

[1] Guillard, Benoit, et al. "Sketch2mesh: Reconstructing and editing 3d shapes from sketches." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021.

AuthorsOpenReview ↗

Author Responses (2)

W2. Design Feature Controllability and the Fidelity–Performance Trade-off.

Sketch anchoring relies on token-wise cosine similarity, so similarity in the sketch latent space does not guarantee preservation of specific design features in the resulting 3D geometry. Moreover, the anchoring strength introduces a trade-off between design fidelity and physical optimization. Additional analysis of these limitations is needed.

A:

  1. Design feature controllability. Design features fall into two kinds: existing features of the initial shape, which should be kept during the optimization, and new features, which should be generated following the target sketch. AnchorGen targets the second kind. We stress that it does not conflict with methods for the first kind, such as GANO [2], which projects each update onto the null space of the constraints on user-specified parts, but complements them. Appendix J (Fig. 18) shows an example: anchored to the sketch of a sedan, AnchorGen turns a wagon into a sedan but moves the locked wheel arches, while GANO keeps the wheel arches but the car remains a wagon. Ours + GANO combines the two: it follows the target sketch while keeping the wheel arches.

  2. Fidelity–performance trade-off. We thank the reviewer for raising this point. We had already analyzed this trade-off in Appendix F (Fig. 7) of the original submission by varying the sketch-to-drag weight ratio, and we will discuss it more prominently in the main text:

    • Anchoring to the shape's own sketch (Fig. 7(b)). As the reviewer notes, the anchoring strength creates a trade-off: a larger weight first keeps more of the source geometry and better stays on the manifold, but too strong an anchor confines the search to a small neighborhood of the source, so drag can no longer descend smoothly. We additionally report, in the new table below, the number of sources (out of 20) that reach the 10% drag reduction.

      Sketch : Drag1:51:31:21:12:13:15:1
      # Success18191816877
    • Anchoring to an exogenous sketch (Fig. 7(a)), whether its shape has higher or lower drag than the source: a larger weight consistently brings the result closer to the target style and, unlike anchoring to the shape's own sketch, does not lower the rate of reaching the drag reduction.

[2] Guoze, Sun, et al. "Geometry-Aware Neural Optimizer for Shape Optimization and Inversion." Forty-third International Conference on Machine Learning.

AuthorsOpenReview ↗

Gentle Reminder: Summary of Our Responses

Dear Reviewer ahar,

Thank you again for your time and valuable feedback. As the discussion period is closing soon, we wanted to gently follow up on our responses above. We believe they address your main concerns:

(W1) Generalization to real freehand sketches, other shape categories, and other physical objectives. We added a chair volume-reduction experiment, where a model trained only on synthetic sketches is tested on real hand-drawn sketches (Sec. 4.4, Appendix H).

(W2) Design feature controllability and the fidelity–performance trade-off. We added a discussion of how different methods preserve design features (Appendix J) and a more detailed analysis of the trade-off across anchoring weights (Appendix F).

If anything remains unclear, we would be happy to discuss further.

Best regards,
AnchorGen Authors

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

AnchorGen addresses physics-driven shape optimization while preserving design intent expressed through a sketch. Its central motivation is that a shape-only generative prior may maintain plausibility while allowing the optimization to change identity-defining design features.

The method learns an unconditional rectified flow over concatenated shape and sketch latents. At inference, it optimizes an intermediate flow state using a differentiable drag surrogate and a sketch-latent anchor, backpropagating through the flow to obtain the resulting shape–sketch pair.

The same model supports preservation of an existing design, editing toward a modified sketch, and sketch-guided generation. Evaluation includes ShapeNet and DrivAerNet++ experiments, with additional OpenFOAM pressure-drag measurements on ten sketch-guided generation cases.

Critical strengths
  • The problem formulation is compelling and practically relevant. Maintaining design intent throughout physical optimization addresses a concrete limitation of workflows in which a sketch influences only the initial geometry.

  • The empirical evaluation includes both shape-only methods and sketch-conditioned alternatives, rather than relying exclusively on baselines without access to design guidance. The 600 source–target pairs on DrivAerNet++ and multiple geometric and dimensional metrics provide substantial evidence of improved target adherence.

  • The paper examines anchor strength, inversion depth, and model capacity, and reports trade-offs on individual geometric metrics. The additional CFD evaluation also provides useful validation beyond improvements measured solely by the optimization surrogate.

Critical weaknesses
  • The claim that optimizing through the flow prevents iterates from leaving the valid-shape manifold is not established. Passing an arbitrarily optimized intermediate state through a learned flow does not guarantee that its output remains in a high-probability region of the data distribution, so the method should be described as providing an empirical prior rather than a manifold-preservation guarantee.

  • Some comparisons confound the optimization mechanism with additional information about the desired solution. In the ShapeNet experiment, target sketches are selected using predicted drag reduction, whereas shape-only baselines do not receive this direction; retrieval-based initialization and simpler optimization methods using the same sketch information would help isolate the proposed mechanism.

  • The implemented update differs materially from the weighted-loss optimization described in the main text. Appendix E.2 and Algorithm 2 introduce gradient normalization, projection onto shape tokens, and Adam updates; moreover, the joint flow can transmit shape-objective gradients to intermediate sketch tokens, so discarding those components requires justification and ablation.

Questions for authors

Could you clarify the source-selection protocol? --- the main text describes randomly selected test vehicles, whereas Appendix B describes feasibility filtering and clustering over the full vehicle pool.

2 comments
AuthorsOpenReview ↗

Author Responses (1)

W1. Manifold-preservation claim.

... Passing an arbitrarily optimized intermediate state through a learned flow does not guarantee that its output remains in a high-probability region of the data distribution,..

A: We thank the reviewer for this valuable remark and agree. We have corrected the imprecise statements in the paper and now describe the method as a better paradigm for injecting a sketch prior to guide shape generation, which is empirically supported by the distributional metrics (Table 1) and pairwise metrics (Table 2) on cars, and the pairwise metrics on chairs (Table 3).

Specifically, we revised the following statements, highlighted in blue in the revised paper:

  • Sec. 1 (l. 094–096): "Every candidate is therefore produced by the joint prior, which empirically keeps it close to the manifold of valid shapes, and the penalty biases it toward the subset consistent with the target."
  • Sec. 3 (l. 268–270): "... every proposed move is re-expressed by the joint prior before it is scored, which empirically keeps the iterate close to the manifold, and the anchor term points the correction to the slice of that manifold selected by the target sketch."

W2. Isolating the Contribution of the Anchor.

... In the ShapeNet experiment, target sketches are selected using predicted drag reduction, whereas shape-only baselines do not receive this direction...

A: We thank the reviewer for this suggestion. Regarding the baselines of Table 1, retrieval-based initialization does not apply: the task there is to optimize a source shape toward a lower-drag design, and initializing from the retrieved shape would not preserve the features of the source shape. More specifically, the tasks in the paper fall into two kinds:

  • Optimizing an existing shape. Starting from the source shape, the goal is to preserve its features as much as possible, add new features, and meet the aerodynamic requirement, e.g., drag reduction for cars.
  • Generating from a design sketch. Starting from scratch, the goal is to preserve the features of the sketch as much as possible while meeting the aerodynamic requirement.

The comparison with PhysGen in Sec. 4.5 (Fig. 6, Table 4) evaluates the second kind, i.e., starting from scratch toward a target sketch, where PhysGen, a larger sketch-conditioned model, receives the same sketch as AnchorGen.

The contribution of the anchor is isolated in two ways, where the information available to all methods is balanced:

  1. No additional target information. In several experiments, the anchor is the sketch of the source shape itself, so it carries no information about a lower-drag solution: Fig. 3 and Fig. 7(b) (Appendix F) on ShapeNet, and Table 3 on chairs, where each chair is anchored to a real hand-drawn sketch of its own original shape. In these settings, AnchorGen still preserves the source design better than the shape-only baselines.
  2. Same sketch information. On DrivAerNet++ (Table 2), all conditioning-based baselines receive the same target sketch as AnchorGen and differ only in how the sketch enters the optimization (Appendix C). The projection baselines in Appendix I (Table 9) also use the same target sketch with a shape-only prior, which corresponds to the simpler optimization methods the reviewer suggests. AnchorGen outperforms both.

Moreover, some baselines in Table 1 also receive explicit drag information: FMG, following FCSO, uses classifier-free guidance conditioned on the target drag value. AnchorGen still outperforms such baselines that are explicitly given the direction of drag reduction.

We have revised the paper to state this distinction more clearly, e.g., in Sec. 4.2 (l. 376–377 and l. 399): "Since shape-only baselines cannot take the target sketch as input, Table 1 measures the benefit of such a design interface rather than comparing methods under equal information; such comparisons are given in Fig. 3, Table 2, and Table 3."

AuthorsOpenReview ↗

Author Responses (2)

W3. Clarification of Implementation Details.

The implemented update differs materially from the weighted-loss optimization described in the main text. Appendix E.2 and Algorithm 2 introduce gradient normalization, projection onto shape tokens, and Adam updates...

A: We thank the reviewer for this careful reading. The gradient normalization in Algorithm 2 addresses two issues of the two terms in Eq. 3: they differ in magnitude, and their gradients may conflict. For example, on ShapeNet, at the start of the optimization the anchor loss is about 0.073 while the drag is about 0.34. With a fixed weight as in Eq. 3, the larger term would dominate the update; normalizing both gradients lets a single ratio control the trade-off.

Gradient normalization and Adam updates are implementation details of the same objective. Regarding the shape-objective gradients that the joint flow transmits to the intermediate sketch tokens, we discard them so that the sketch tokens focus on preserving the style of the target sketch, and the drag objective does not interfere with them and slow down convergence. We provide an additional ablation in the setting of Table 1, where the metrics are computed on the 92 cases in which both variants reach the 30% drag reduction:

MethodSuccess / 100FD_incep ↓KD_incep ↓FD_dino ↓KD_dino ↓FD_point ↓KD_point ↓
w/o projection9276.4243.4457.9043.4515.2430.16
AnchorGen (Ours)9876.9043.5257.9543.6815.8732.88

Without the projection, the metrics are similar, but more iterations are needed and fewer cases reach the target (92 vs. 98).

Q1. Source-selection protocol.

Could you clarify the source-selection protocol? --- the main text describes randomly selected test vehicles, whereas Appendix B describes feasibility filtering and clustering over the full vehicle pool.

A: We apologize for the unclear description. Each sketch shows a single view and thus prescribes a change in a single dimension, e.g., the length in the side view. Since the pairwise metrics compare the result with the full 3D target, differences in the other dimensions would count as errors that the sketch could not prescribe. We therefore select each source/target pair to differ in a single dimension (length, width, or roof height) while keeping the other two as similar as possible.

The source shapes, which are the starting points of the optimization, are all taken from the held-out splits and are unseen during training. Appendix B refers to the full pool of vehicles only because the standardization statistics of the three dimensions are computed over the whole pool, and the target sketches that increase or decrease each dimension are matched from it. To ensure that every source can be matched with both a larger and a smaller target in each dimension, we restrict the source pool to vehicles that admit all six edits, which excludes vehicles at the edge of the distribution. Among the 274 such vehicles outside the training split, we select 100 sources by kk-means, taking from each cluster the vehicle with the largest feasible edits.

We have revised the main text and Appendix B accordingly:

  • Sec. 4.3 (l. 412–415): "We select 100 source vehicles outside the training split and pair each with six targets, obtained by stretching or compressing one of length (side view), width (front view) or height (rear view) while keeping the other two dimensions nearly unchanged (Appendix B) ..."
  • Appendix B (l. 958–960): "Of the vehicles, 987 admit all six requested edits, which excludes vehicles at the edge of the distribution; we select the 100 sources only among the 274 of them outside the training split ..."

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers,

We sincerely thank all reviewers for their valuable comments and insightful suggestions. We are encouraged that the reviewers recognize the practical relevance of preserving design intent during physics-driven optimization (EvDz, 2u5r), the flexibility of combining design constraints and physical objectives without retraining (ahar), and the thorough evaluation against both prior-based and conditioning-based baselines (EvDz, ahar). We are also pleased that Reviewer 2u5r finds the paper clear and easy to follow.

We have carefully addressed all the reviewers' concerns in the comments and revised the paper. For clarity, the revised parts of the manuscript and appendix are highlighted in blue. The primary changes are summarized as follows:

  1. Added a chair volume-reduction experiment with real hand-drawn sketches (Sec. 4.4, Appendix H).
  2. Added an ablation comparing joint modeling with sketch-projection consistency losses (Appendix I).
  3. Clarified the scope of our claim on design features: AnchorGen generates new features following the target sketch and complements methods that preserve existing features, such as GANO (Appendix J).
  4. Added compliance minimization on 21 ShapeNet categories with a single model, under category-specific load cases (Appendix K).

Best regards,

AnchorGen Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.