ViewCraft3D: High-Fidelity and View-Consistent 3D Vector Graphics Synthesis

TL;DR

ViewCraft3D (VC3D) tackles the challenge of generating high-fidelity, view-consistent 3D vector graphics from a single image. It leverages 3D priors and a two-stage optimization: Stage I fits cubic Bézier curves to salient 3D structure extracted from a mesh reconstructed by an image-to-3D model, and Stage II refines missing details with Score Distillation Sampling guided by a pretrained model. The method renders 3D curves to 2D views via differentiable projection and demonstrates superior view consistency and aesthetics while reducing generation time compared with prior 3D vector graphics approaches. This work improves accessibility to expressive 3D vector graphics for VR, shape retrieval, and conceptual design by avoiding heavy 2D priors and cross-view inconsistencies.

Abstract

3D vector graphics play a crucial role in various applications including 3D shape retrieval, conceptual design, and virtual reality interactions due to their ability to capture essential structural information with minimal representation. While recent approaches have shown promise in generating 3D vector graphics, they often suffer from lengthy processing times and struggle to maintain view consistency. To address these limitations, we propose ViewCraft3D (VC3D), an efficient method that leverages 3D priors to generate 3D vector graphics. Specifically, our approach begins with 3D object analysis, employs a geometric extraction algorithm to fit 3D vector graphics to the underlying structure, and applies view-consistent refinement process to enhance visual quality. Our comprehensive experiments demonstrate that VC3D outperforms previous methods in both qualitative and quantitative evaluations, while significantly reducing computational overhead. The resulting 3D sketches maintain view consistency and effectively capture the essential characteristics of the original objects.

Figures (8)

• Figure 1: We propose ViewCraft3D (VC3D), a method to generate 3D vector graphics from a single image. VC3D can leverage 3D prior knowledge to generate high-quality and view-consistent 3D vector graphics.
• Figure 2: Examples of VR sketches Sketchfab.
• Figure 3: The overall architecture of the proposed method, showcasing the initial generation of 3D Vector Graphic (3D VG) from an input image and subsequent detail refinement using a pretrained image-to-3D model.
• Figure 4: The visualization process of Point Cloud Clustering. Each point is assigned an orientation vector (blue arrows). In (a), the orange point initializes the cluster, and a candidate point (red) is evaluated based on spatial proximity and orientation similarity. In (b), the candidate point meets both criteria and is incorporated into the cluster. This process iterates until the final cluster is formed, as shown in (d).
• Figure 5: Qualitative comparison of different methods. Diff3DS and VC3D use a single image $\bm{I}$ as input, while 3Doodle uses $120$ rendered images of the mesh reconstruction result $\mathcal{M}$ as input.