M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals

arXiv cs.GR details M-plicits, a neural implicit surface framework built around nested multiscale residual MLPs. The pitch is straightforward: keep the flexibility of implicit surfaces, but avoid the usual tradeoff between inference cost, noisy reconstructions, and brittle multiscale artifacts.

The key idea is to train a coarse network first, then add residual networks in progressively tighter neighborhoods around the previous zero-level set. That local supervision acts like a geometric low-pass filter, so the early stage establishes a cleaner shape prior while later stages refine detail without chasing high-frequency noise.

For developers, the practical angle is in the rendering path. M-plicits pairs the reconstruction method with a multiscale sphere-tracing algorithm and GEMM-based analytic normal computation, sidestepping auto-differentiation for high-fidelity real-time rendering. The authors say this also avoids the costly mesh extraction step often needed for visualization.

On Stanford and Thingi32, the method is reported to achieve the best mean Chamfer distance in the coarse setting and the best median Chamfer distance and IoU in the fine setting, while using about an order of magnitude fewer parameters than grid-based baselines. The paper also claims stronger noise robustness than iNGP, BACON, and IDF, which makes it especially relevant for reconstruction pipelines that have to deal with imperfect scans or captured geometry.