3D target object generation based on spatial consistency

• 1、School of Intelligent Engineering and Automation, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Text-to-3D generation via score distillation has achieved remarkable progress by leveraging powerful 2D diffusion priors. However, the Janus problem---where generated 3D objects exhibit duplicated or geometrically inconsistent structures across different viewpoints---remains a fundamental challenge. Recent methods such as RecDreamer introduce Uniform Score Distillation (USD) with an orientation classifier to encourage pose-uniform generation, yet rely on manually selected DINOv2 templates that are prompt-specific, labor-intensive, and prone to front-view bias on NeRF-rendered images. In this work, we propose Sco3D, a framework that addresses the Janus problem through two complementary discriminators. First, we introduce an Orientation-Aware Discriminator built upon Orient-Anything, a zero-shot azimuth estimator that replaces the template-based classifier with a fully automatic, prompt-agnostic 360$^\circ$ orientation predictor. Second, we propose an Instance-Consistent Discriminator that leverages DINOv2 CLS token features to enforce multi-view identity coherence, ensuring that renderings from different viewpoints correspond to the same underlying entity. Both modules are seamlessly integrated into the USD training pipeline as plug-and-play components. Extensive experiments demonstrate that Sco3D significantly reduces Janus artifacts while maintaining high-fidelity text-3D alignment, outperforming existing score distillation methods across both qualitative and quantitative evaluations.

Keywords： Computer Vision and Application；3D Generation；Spatial Sonsistency