MeshFM: 2D Features Are All You Need for 3D Shape Understanding

MeshFM is a new feedforward framework for extracting rich 3D shape features by distilling information from 2D visual foundation models. Instead of relying on expensive 3D annotations or iterative optimization at inference, it predicts 3D features directly, which makes the pipeline much more practical for real production workflows.

The training setup is split into two stages. First, a 3D feature field is optimized using only 2D feature supervision. Then a network is trained to regress that field, so the final model can run without per-shape fitting at inference time. That matters for teams that need scalable preprocessing or interactive tools, where optimization-heavy methods can become a bottleneck.

The claimed payoff is broad downstream usefulness: the learned features transfer immediately to part segmentation, dense correspondence, and mesh deformation. The model is reported to perform on par with methods trained explicitly with 3D supervision, even without task-specific fine-tuning, and it is designed to handle extreme rotations of input objects more robustly than many shape pipelines.

For game developers, the interesting angle is practical asset understanding. Better automatic mesh features could help with rigging, deformation tools, semantic tagging, content search, and procedural workflows, especially when studios have lots of meshes but limited labeled 3D data. The exact integration cost will depend on implementation details, but the direction is clear: 2D priors are becoming strong enough to bootstrap useful 3D tooling.