Turning a drone’s worth of aerial photos into a usable, lightweight 3D building model — a “textured proxy” — is standard practice in game environments, digital twins, and procedural city generation. The standard pipeline is strictly sequential: reconstruct the geometry first, from a dense point cloud or mesh, then project textures from the photos onto that shape afterward. A paper accepted to SIGGRAPH Asia 2026, posted to arXiv this month, argues that order is the whole problem.
Where the sequential approach breaks
Once geometry is “locked in” from the dense reconstruction alone, any mismatch between where the photos say an edge is and where the simplified geometric proxy puts it has nowhere to go but the texturing stage — where, by definition, you can no longer move the geometry to fix it. The authors (Weidan Xiong, Shihao Feng, Kaijie Qu, Hao Li, Mingqi Zhang, Daniel Cohen-Or, and Hui Huang) point to exactly the failure modes you’d expect from that constraint: facades and their intersecting edges drift out of alignment across different camera viewpoints, producing duplicated edges, texture drift, and visibly inconsistent seams between views in the final model.
Building geometry and texture together
The paper’s fix is to make proxy generation image-guided from the start: the geometry optimization is constrained jointly by the dense reconstruction and structural cues pulled directly from the input aerial images — facade lines, edges, and where they should intersect — rather than letting the geometry solidify independently and hoping the textures cooperate later. Structural priors extracted from the images and from the dense geometry are enforced to stay consistent with each other throughout the optimization, not just checked at the end.
The result, per the authors’ own experiments across a range of architectural scenes, is a compact, geometry-texture-aligned proxy with substantially fewer of the edge-duplication and cross-view inconsistency artifacts that plague the sequential approach — without needing a larger model or a separate correction pass bolted onto the pipeline afterward.
Why this matters beyond architecture
For creative coders and procedural-generation work specifically, this is a reminder that the order of operations in a generative pipeline — not just the individual algorithms — can be the actual source of a persistent quality ceiling. A technique built for photogrammetry-to-proxy workflows has an obvious downstream use in procedural city generation and digital-twin work, anywhere a lightweight, texture-accurate building stand-in needs to hold up under motion and close inspection rather than just look right in a single still render.
The paper carries no project page or public code repository as of this writing — it’s a fresh November submission (arXiv 2611.12077) heading toward SIGGRAPH Asia 2026 in Kuala Lumpur, which this site previewed earlier this week.