Generating high-fidelity, city-scale 3D environments is essential for applications like virtual reality, gaming, and urban digital twins. Yet the field is held back by a fundamental bottleneck: real-world city-scale 3D training data is scarce, making it difficult to learn direct texturing methods.
The arXiv paper "CityPaint: A Free Lunch from 2D Diffusion for City-Scale 3D Texturing" proposes a way around this limitation. Instead of relying on large 3D datasets, the method appears to repurpose powerful 2D diffusion models to apply textures to 3D city models, treating them as a readily available resource.
The title's "free lunch" suggests the approach extracts 3D texturing capability from existing 2D models without additional training data. While the abstract is brief, the implication is that this could lower the barrier for creating detailed virtual cities, though the paper's full methodology and results remain to be examined.