ASV3D: Adapting Diffusion-Based Single-View 3D Reconstruction with Extra Imagery

arXiv cs.GR details ASV3D, a framework for adapting diffusion-based single-view 3D reconstruction with support from one additional image at test time. The core idea is simple but useful: instead of forcing a model to infer missing geometry from one view alone, ASV3D folds in an auxiliary image to recover more complete shapes.

The paper describes two modes. A zero-shot adaptation path uses the extra view without retraining, while an optimized variant adds contrastive learning to push visual fidelity and cross-view consistency further. In practice, that makes the system more robust when inputs are unconstrained and viewpoints are messy, which is exactly where single-image reconstruction tends to fall apart.

The authors say ASV3D improves two state-of-the-art diffusion-based reconstruction pipelines on benchmark and real-world datasets, with gains in quantitative metrics and human preference. They also published code and a real-world object dataset, which should make the work easier to reproduce and adapt for research or production prototyping.

For game developers, the interesting part is the workflow implication: if you’re building tools for photogrammetry, asset ingestion, or rapid object capture, a lightweight second view can buy noticeably better geometry without a full retraining pass. That makes the approach relevant to technical artists and tools teams looking for practical reconstruction quality gains.