HiFiVe: High-Fidelity Vehicle Generation Leveraging Auto-Regressive 2D Generative Priors

Existing 3D vehicle generation methods often produce models with low geometric fidelity and blurry textures. While multi-view diffusion models have improved texture quality, they typically require costly fine-tuning, are limited to fixed viewpoints, and struggle with cross-view consistency. This paper introduces HiFiVe, a training-free framework designed to enhance both texture and geometry for 3D vehicle modeling.

HiFiVe operates in three main stages. First, it employs an autoregressive texture refinement pipeline that synthesizes high-resolution textures sequentially. To maintain consistency, it uses the coarse mesh as a geometric synchronization prior, warping previously generated frames into the current viewpoint using depth-based projection. A multi-view texture fusion strategy assigns higher confidence to pixels that align with surface normals, effectively mitigating ghosting artifacts.

Second, the authors optimize the generation sequence by exploiting the inherent bilateral symmetry of vehicles. By using a BFS-style generation path, the model minimizes the autoregressive distance from the initial anchor view, which suppresses error accumulation over long sequences. Finally, the framework recovers high-frequency surface details by refining the mesh geometry using normal maps estimated from the enhanced textures. A frequency-adaptive weighting strategy is applied during this optimization to ensure that high-frequency geometric features are prioritized while maintaining global structural stability.

Experiments on both synthetic (SketchFab-Cars) and real-world (3DRealCar) datasets demonstrate that HiFiVe significantly outperforms state-of-the-art baselines in both visual texture fidelity and geometric detail. By leveraging powerful 2D generative priors (Flux.2) without the need for additional training, the framework provides a flexible and efficient solution for creating high-quality 3D vehicle assets suitable for applications like autonomous driving simulation and virtual content creation.