Do AI vision models understand space when images are blurry or dark?

Current multimodal AI models excel at spatial reasoning on clean images but fail dramatically when faced with real-world visual degradation: blur, low light, weather artifacts, and compression. SpaceDG introduces a 1M-question dataset with physically realistic degradations synthesized through 3D Gaussian Splatting, plus a verified benchmark of 10K test cases. Testing 25 models exposed consistent robustness gaps across degradation types—but the key finding: training on degraded images restores spatial intelligence and even surpasses human performance without hurting clean-image accuracy.