M-plicits
A neural implicit surface reconstruction method based on a nested multi-scale residual architecture. It employs a hierarchical residual network to progressively refine geometric details atop a coarse surface baseline, balancing robustness against noise with the preservation of high-frequency geometric detail.
Timeline (last 90 days)
M-plicits was accepted at NeurIPS 2026.
M-plicits reconstructs neural implicit surfaces by first learning a clean coarse surface, then progressively refining it using residual networks within nested neighborhoods around the previous layer's surface.
M-plicits decomposes a complex surface into a coarse base plus multiple layers of residual detail, with each residual layer trained within a nested neighborhood near the previous layer's surface.
M-plicits unifies multiple levels of detail into a single representation, organized into a hierarchical structure through nested residuals, enabling controllable noise/detail trade-offs.
The authors provide an intuitive comparison between M-plicits and iNGP (Instant NGP) on noisy inputs on the project page.
M-plicits proposes companion algorithms including multi-scale sphere tracing, analytic surface normals, surface extraction, and neural normal mapping.
The multi-scale structure of M-plicits allows surface normals to be derived analytically, without relying on automatic differentiation.
The authors have publicly released the code, experimental data, and 105 checkpoints covering 36 shapes with training configurations, with models hosted on Hugging Face.
M-plicits checkpoints are individually fitted independent surfaces, with each shape trained separately; fitting a new shape requires retraining, and the method does not possess cross-shape generalization capability.
As the first author, the researcher invited the community on Reddit to provide feedback on whether the noise/detail trade-off is reasonable and on cross-method comparisons with other implicit surface representations.
All Facts (10)
M-plicits reconstructs neural implicit surfaces by first learning a clean coarse surface, then progressively refining it using residual networks within nested neighborhoods around the previous layer's surface.
50%UnverifiedM-plicits checkpoints are individually fitted independent surfaces, with each shape trained separately; fitting a new shape requires retraining, and the method does not possess cross-shape generalization capability.
50%UnverifiedM-plicits decomposes a complex surface into a coarse base plus multiple layers of residual detail, with each residual layer trained within a nested neighborhood near the previous layer's surface.
50%UnverifiedM-plicits unifies multiple levels of detail into a single representation, organized into a hierarchical structure through nested residuals, enabling controllable noise/detail trade-offs.
50%UnverifiedThe authors provide an intuitive comparison between M-plicits and iNGP (Instant NGP) on noisy inputs on the project page.
50%UnverifiedM-plicits proposes companion algorithms including multi-scale sphere tracing, analytic surface normals, surface extraction, and neural normal mapping.
50%UnverifiedThe multi-scale structure of M-plicits allows surface normals to be derived analytically, without relying on automatic differentiation.
50%UnverifiedAs the first author, the researcher invited the community on Reddit to provide feedback on whether the noise/detail trade-off is reasonable and on cross-method comparisons with other implicit surface representations.
50%UnverifiedThe authors have publicly released the code, experimental data, and 105 checkpoints covering 36 shapes with training configurations, with models hosted on Hugging Face.
50%UnverifiedM-plicits was accepted at NeurIPS 2026.
50%