When AI Augmentation Helps 3D Gaussian Splatting: Perceptual Thresholds for Training Views

Image-based 3D reconstruction and visualization, such as 3D Gaussian Splatting (3DGS), enable high-fidelity representations for free-viewpoint navigation. These methods typically require dense captured images for effective training, and reconstruction quality degrades when samples are limited. Recent work has explored using generative AI to synthesize additional training views under limited-data conditions, but the perceptual impact of AI-generate dimages in 3DGS pipelines remains unclear. We conduct a perceptual user study of 3DGS visual quality across varying data conditions. Our results show that AI-generated views can improve perceived quality when real samples are sparse or unevenly distributed, but the benefits plateau and can degrade beyond certain thresholds. We identify these perceptual transition points and discuss implications for data-efficient 3D reconstruction in immersive applications.