W1: Novelty.

While our pipeline leverages pretrained models, our core contribution is the introduction of a novel framework that directly addresses the fundamental issue of view-inconsistency prevalent in existing 3D graphics generation methods. Previous works, such as 3Doodle[1], Diff3DS[2], and Dream3DVG[3], rely heavily on 2D constraints: they project 3D models into 2D views and utilize differentiable rendering to enforce alignment between the projections and image ground-truth. However, this reliance on 2D priors inherently limits their ability to enforce true 3D coherence across multiple views.

In contrast, our method introduces a paradigm shift by incorporating 3D priors directly into the optimization process for 3D vector graphics. Specifically, we extract salient edges from the mesh and leverage a pretrained 3D generative model to define a 3D SDS loss in latent space, thereby completely removing the dependency on 2D projections. This enables optimization to be guided by a holistic understanding of 3D shape geometry, rather than by enforcing consistency across 2D views.

To the best of our knowledge, ViewCraft3D is the first framework to apply 3D SDS loss for optimizing structured 3D vector graphics, which introduces both conceptual and technical novelty beyond a simple combination of existing components.

Q1: Bézier Fitting.

Number of Curves: The number of curves is jointly determined by the number of clusters, resulting from the Point Cloud Clustering stage (Section 3.2.2), and the filtering threshold, $\tau$. The filtering threshold removes clusters with fewer than $\tau$ points, and each of the remaining clusters then has a Bézier curve fitted to it.

Region Identification and Curve Initialization: As we detailed in Lines 204-214, regions for refinement are identified by analyzing the mesh's vertex distribution, locating areas not adequately covered by the point cloud $\mathcal{P}_c$ from Stage I (Section 3.2). Specifically, a vertex is considered "uncovered" if there are no points from point cloud $\mathcal{P}_c$ within a radius $R$ of that vertex. To initialize new Bézier curves, we first apply K-Means clustering to uniformly sample points from the uncovered point set, using them as initial control points. The remaining control points are then randomly selected within a confined local region around each initial point, forming the initial curves.

Q2: Complexity.

Thank you for the question. Our method may produce more curves. However, this reflects the model’s ability to capture finer geometric details. To quantitatively investigate the complexity–quality trade-off, we conducted an experiment using ten relatively complex examples selected from those reported in the main paper. The quantitative results are presented in the table below.

We can see with the filtering threshold τ increased, the CLIP score and Aesthetic score decrease, suggesting that geometric details are gradually reduced.

Q3: Details of Diffusion Model.

The goal of the refinement stage is to improve the initial Bézier curve representation obtained in Stage I by leveraging 3D generative priors via Score Distillation Sampling (SDS). SDS aims to generate a pseudo-supervisory signal from a pretrained diffusion model (in our case, an image-conditioned 3D diffusion model) and iteratively adjusts the target (here, the Bézier curve parameters $\theta$) via gradient descent.

Our method builds upon the architecture of TripoSG, which comprises two main components: 3D Variational Autoencoder (VAE) and a Rectified Flow Transformer. In our pipeline, we first sample a point cloud from the Bézier curves and encode it into the latent space $\mathcal{Z}$ using the pretrained VAE encoder from TripoSG, obtaining a latent code $z$.

We then apply SDS by perturbing this latent code and feeding it into the Rectified Flow Transformer, which returns a score (i.e., a denoising direction). This score acts as a gradient signal that suggests how the latent code—and consequently the underlying 3D curves—should be modified to better match the priors captured by the diffusion model. We backpropagate this gradient through the VAE encoder and the point sampling process to update the Bézier curve parameters $\theta$, thereby refining the 3D curves in a differentiable, geometry-aware fashion.

[1] 3Doodle: Compact Abstraction of Objects with 3D Strokes

[2] Diff3DS: Generating View-Consistent 3D Sketch via Differentiable Curve Rendering

[3] Empowering Vector Graphics with Consistently Arbitrary Viewing and View-dependent Visibility

Dear Reviewer 221W,

Thank you again for your valuable feedback on our paper.

We have submitted our rebuttal, which addresses the concerns you raised, including our novelty, algorithm details and complexity–quality trade-off.

As the discussion period concludes in two days, we would be grateful if you could take a moment to review our response and let us know if you have any further questions.

W1: The Role of Reconstructed Mesh.

Thank you for your question. The most critical idea of our approach is to achieve view consistency by leveraging 3D priors for 3D vector graphics generation. The 3D priors include geometric priors from the reconstructed mesh (Stage I) and generative priors from a pretrained 3D generative model (Stage II). Therefore, the quality of the generated 3D vector graphics is not simply determined by the quality of the mesh.

Specifically, in Stage I, we extract only the most salient edges from the reconstructed mesh, intentionally filtering out noise and surface details to mitigate the influence of potential artifacts or inaccuracies in the mesh. In Stage II, the vector graphic is further refined via SDS optimization, using 3D generative priors learned from large-scale 3D assets.

As a result, view consistency achieved by our method is not passively inherited from the mesh, but is actively enforced through our two-stage pipeline and the integration of 3D priors.

W2 & Q1: Application of 3D Vector Graphics.

We appreciate the reviewer’s concern regarding the utility of 3D vector graphics beyond aesthetic purposes. Our method produces a lightweight view-consistent 3D vector representation that offers distinct advantages for downstream tasks where interpretability, interactivity, and performance efficiency are critical.

For example, in an AR-based industrial assistance application, the generated vector graphic can be overlaid on real-world machinery to provide intuitive visual guidance. Instead of rendering a full textured 3D mesh, which is often computationally expensive, the vector abstraction can highlight the outline of a component that needs replacement or trace the routing of a cable, enabling fast, responsive, and informative AR interactions, especially on resource-constrained devices.

Beyond AR, our representations are also valuable in CAD and industrial design, where curve-based editing and structural manipulation are standard. Designers can interactively modify part outlines, reuse components, or perform topological edits without operating on dense, hard-to-parse mesh data. Similarly, in education and training, 3D vector graphics allow learners to better understand object structures through simplified line-based visualizations, which can emphasize functionally important parts over surface-level details.

Thus, our use of 3D vector graphics is not merely for aesthetics, but reflects a deliberate design trade-off that prioritizes editability, abstraction, and semantic clarity. Such capabilities are not inherently provided by raw 3D meshes. For more specific application scenarios, please refer to our response to Reviewer ztRZ W3.

Q2: Occlusion Limitation.

Sorry for the confusion. The occlusion limitation mentioned in the paper specifically refers to the 2D rendering of the 3D vector graphic for visualization purposes, and is not related to the quality of the generated 3D curve representation itself. Since our current rendering assumes the same opacity and no depth ordering (i.e., all curves are drawn with the same style regardless of their spatial relationships), this can occasionally lead to reduced visual clarity and spatial interpretability, especially in cases involving dense overlaps or self-occluding structures. We will include qualitative examples of failure cases in the revised version.

We appreciate the reviewer's thoughtful engagement with our work. We believe the additional experiments and clarifications we have provided should address the concerns raised. We welcome any further questions or comments during the remaining discussion period and are committed to providing comprehensive responses.

W1: Texture Synthesis.

Thank you for this insightful comment. The absence of texture is an intentional design choice, as our paper's core objective is to generate view-consistent 3D vector graphics by leveraging 3D geometric priors. Our method focuses exclusively on these geometric features to solve the central challenge of view consistency for an object's structure. We agree that incorporating texture is an important direction, and we will explore it in our future work.

W2 & Q1: Clustering Method and Ability of Producing Long Strokes.

Our clustering method is capable of generating long and continuous strokes, as demonstrated by the global outlines of the coffee cup and scissors in Figure 5. Unlike 2D abstraction approaches such as CLIPasso, which are optimized for perceptual minimalism, our task requires balancing coherence across multiple viewpoints with localized geometric detail. Therefore, there is a trade-off between stroke length/number and detail preservation.

Specifically, for objects with rich details, lowering hyperparameter $\tau$ retains more clusters, yielding more Bézier curves and richer details. For objects best described by long strokes, increasing hyperparameter $d_{thresh}$ encourages each cluster to extend longer, thereby yielding longer Bézier curve fitting results.

W3 & Q2: The Role of Primary Structure Fitting.

Thank you for your question. Primary Structure Fitting is a critical component that significantly enhances both generation quality and efficiency. To demonstrate its importance, we conduct an ablation study by removing Primary Structure Fitting (Stage I) and relying solely on SDS loss for optimization. The results are shown in the following table:

We can see that the CLIPScore decreases from 0.799 to 0.622, and the Aesthetic Score drops from 4.352 to 3.653. Additionally, we perform qualitative comparison and observe a noticeable degradation in 3D view consistency compared to our full method. The visualization results show that the lines in SDS-only generated results are more cluttered and lack sufficient structural lines. Moreover, the SDS-only optimization is extremely inefficient, requiring approximately 3 hours to generate a vector graphic. Similar results are also observed in related works like DiffSketcher[1] (Figure 8). Therefore, Primary Structure Fitting is an indispensable step that provides the necessary structural foundation for the refinement stage. We will include this ablation study in the next version.

W4: Potential Contradict.

We appreciate the reviewer’s insightful observation regarding the potential contradiction between the point cloud prior used in SDS and the salient edge-based curve fitting in Stage I.

As we expalined in Line209-211 of our submission, the final 3D SVG composite of two sets of Bezier curves, curves from Stage I and II. Specifically, curves from Stage I fit to salient edges, capturing the primary structural features of the mesh. At Stage II, we initialize new curves in regions that are geometrically complex but not well approximated by Stage I, such as flat or smooth surfaces with fine-scale details. This design ensures that the aggregated point cloud better aligns with the distribution expected by the pretrained model used in SDS, which is based on full-surface point clouds.

Moreover, our empirical results show that the TripoSG encoder is still capable of encoding the Bézier curve-sampled point cloud into valid latents. By applying noise to the latent code and denoising through the Rectified Flow model, we can reconstruct the overall 3D shape effectively. This demonstrates that, although our point cloud is sampled differently, it still falls within the distribution that TripoSG can effectively encode.

Q3: Adapt to Other 3D Generation Frameworks.

Our VC3D can be adapted to other 3D generative frameworks beyond TripoSG. The key requirement is that the framework must provide an encoder. This encoder is essential for our method, as it takes the point cloud sampled from the 3D Bézier curve representation and encodes it into latent space, which is then used for optimization via Score Distillation Sampling (SDS). While other frameworks, such as Hunyuan3D[2], are potential candidates, many have only released the decoder code. Therefore, we selected TripoSG as it offers a publicly available encoder that fulfills this requirement and also demonstrates strong generation performance.

[1] DiffSketcher: Text Guided Vector Sketch Synthesis through Latent Diffusion Models.

[2] Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation

Dear Reviewer X23k,

Thank you again for your valuable feedback on our paper.

We have submitted our rebuttal, which addresses the concerns you raised, including long strokes and the role of Primary Structure Fitting.

As the discussion period concludes in two days, we would be grateful if you could take a moment to review our response and let us know if you have any further questions.

W1 & Q1: Novelty.

Our core contribution is the introduction of a novel framework that directly addresses the fundamental issue of view-inconsistency prevalent in existing 3D graphics generation methods. Previous works, such as 3Doodle[1], Diff3DS[2], and Dream3DVG[3], rely heavily on 2D constraints: they project 3D gaphics into 2D views and utilize differentiable rendering to enforce alignment between the projections and image ground-truth. However, this reliance on 2D priors inherently limits their ability to enforce true 3D coherence across multiple views.

In contrast, our method introduces a paradigm shift by incorporating 3D priors directly into the optimization process for 3D vector graphics. Specifically, we extract salient edges from the mesh and leverage a pretrained 3D generative model to define a 3D SDS loss in latent space, thereby completely removing the dependency on 2D projections. This enables optimization to be guided by a holistic understanding of 3D shape geometry, rather than by enforcing consistency across 2D views.

To the best of our knowledge, ViewCraft3D is the first framework to apply 3D SDS loss for optimizing structured 3D vector graphics, which introduces both conceptual and technical novelty beyond a simple combination of existing components.

W2 & Q2: Comparison with Dream3DVG.

We should clarify that Dream3DVG is concurrent work, since its preprint was released (May 27) after the NeurIPS submission deadline. Therefore, the absence of this comparison in our initial submission is not a weakness. Nevertheless, we have conducted a quantitative comparison. Specifically, given that Dream3DVG accepts text prompts as input, we manually convert the input images from our quantitative experiments into text prompts and use them as input for Dream3DVG. As shown in the following table, our method achieves a higher aesthetic score than Dream3DVG.

In additional to the quantitative analysis, we also conducted a qualitative comparison between our method and Dream3DVG on two representative categories of objects. Given the format restrictions that no addtional images can be added during the rebuttal phase, we describe the results below:

(1) For man-made objects with strong symmetry and straight-line dominance, such as cars, pistols, and houses, Dream3DVG often produces results resembling a 3D version of CLIPasso. It tends to omit or oversimplify parallel structural lines representing object thickness. For instance, vertical lines indicating the walls of a house are frequently missing. We hypothesize this is due to misleading optimization signals: parallel structural lines that should be generated simultaneously may appear overlapping or closely spaced in certain viewpoints, causing the optimization process to favor only one of them and ignore the rest.

(2) For natural objects with asymmetric and richly curved structures, such as birds, butterflies, and flowers, Dream3DVG suffers from more pronounced view inconsistency issues. Taking the bird example (our result shown in Figure 1 of the main paper), Dream3DVG generates duplicated features from different viewpoints. For example, the beak and tail feathers appear twice, and an extra eye is hallucinated on the back of the head. These results demonstrate Dream3DVG's challenge in maintaining coherent and unified vector abstraction across diverse views of natural objects.

These comparisons suggest that while Dream3DVG performs reasonably well on simple and rigid structures, it struggles with both the geometric completeness required for structured man-made objects and the multi-view consistency critical for deformable, natural categories. We will include the above results and discussion in the revised version.

W3 & Q3: Applications of Our Method and Resultant 3D SVGs.

The inherent editability of 3D vector graphics offers significant advantages for iterative modeling workflows, a potential explored in several academic works. (1) 3D vector curves can serve as real-time guidance for surface creation and editing. For instance, FiberMesh[4] enables the design of freeform surfaces by allowing users to directly draw and manipulate a network of 3D curves, which function as a structural vector scaffold for the surface. (2) Some works focus on reconstructing 3D surfaces from a network of 3D vector curves, treating the curve network as an editable and interpretable representation that facilitates intuitive shape refinement. Notably, True2Form[5] infers a complete 3D curve network from ambiguous 2D sketches, bridging the gap between sparse 2D input and structured 3D geometry.

3D vector graphics also support animation production. For example, Blender provides an official SVG plugin and keyframe interpolation tools. Users can project 3D vector models onto different viewpoints, and then generate animations using the plugin. This enables the creation of 2D animations from multiple angles while maintaining strong frame-to-frame consistency.

3D vector graphics are integral to VR/AR applications, and they have already been applied in some well-known software, such as Tilt Brush, Gravity Sketch, and Mental Canvas. With the rapid development of VR/AR devices and applications, 3D vector graphics will become more valuable. Our proposed approach can be used to generate 3D vector graphics assets.

[1] 3Doodle: Compact Abstraction of Objects with 3D Strokes

[2] Diff3DS: Generating View-Consistent 3D Sketch via Differentiable Curve Rendering

[3] Empowering Vector Graphics with Consistently Arbitrary Viewing and View-dependent Visibility

[4] FiberMesh: Designing Freeform Surfaces with 3D Curves

[5] True2Form: 3D curve networks from 2D sketches via selective regularization

We appreciate the reviewer's thoughtful engagement with our work. We believe the additional experiments and clarifications we have provided should address the concerns raised. We welcome any further questions or comments during the remaining discussion period and are committed to providing comprehensive responses.