Anyone who has written a raymarched shader knows the appeal of implicit surfaces: you don’t store geometry, you write a function that says how far away the surface is, and the shape falls out. Signed distance fields are the backbone of a huge amount of procedural and generative 3D work.
Learning those functions from data — implicit neural representations — has become the standard way to turn a scan or a point cloud into a smooth, resolution-independent surface. The technique works, and it forces an annoying three-way trade.
M-plicits, posted 23 September 2026, is an attempt to stop paying for all three at once. The authors are Vinícius da Silva, Isabelle Melo, Matheus Bessa, Guilherme Schardong, Luiz Schirmer, André Araújo and Nuno Gonçalves.
The trade-off being attacked
Encoding input coordinates with sinusoidal functions into MLPs works well for surfaces defined as zero-level sets. But, as the paper lays out:
- Single-MLP approaches are expensive at inference — every query runs the whole network
- Grid-based representations are fast but can limit surface smoothness and overfit input noise
- Previous multiscale approaches frequently capture noise and produce artifacts, because of hard spectral truncation
That third one is the specific target. Multiscale is the obvious idea — coarse network for overall shape, finer networks for detail — but cutting between frequency bands abruptly introduces its own artifacts.
What M-plicits does
Two ideas, and the second is the one that matters.
A residual sum of MLPs. The surface is modelled as several networks summed together, each trained in sequence rather than one large network doing everything.
Nested neighbourhoods with strictly localised supervision. This is the departure. Existing residual approaches sample across the whole domain and need costly mesh extraction just to visualise intermediate results. M-plicits instead confines each stage’s supervision to narrow bands around the previous stage’s zero-level set.
So stage one learns a rough surface. Stage two only looks at a thin shell around where stage one thinks the surface is, and learns a correction there. Stage three narrows further.
Why this fixes the noise problem almost for free
The nested design naturally provides robustness against noisy input data, and the mechanism is elegant: the coarse network acts as a low-pass filter.
A coarse MLP trained on the whole domain cannot represent high-frequency detail, so it cannot represent high-frequency noise either — it fits the underlying shape and ignores the jitter. Subsequent stages then only ever see a narrow band around that already-smoothed estimate. The noise in the original data mostly falls outside the band being supervised.
You get denoising as a property of the architecture rather than as a preprocessing step with parameters to tune. For anyone who has fought a scan with a bilateral filter and a guessed radius, that’s the appealing part.
Confining supervision to narrow bands also means less wasted computation on empty space, and it removes the need for mesh extraction to inspect intermediate stages.
What a creative coder would do with it
Three honest use cases, and one caveat.
Photogrammetry and scan cleanup. If your pipeline is phone scan → point cloud → mesh → clean up by hand, an architecture that filters noise structurally is worth watching. This is the most direct application.
Level-of-detail for free. A residual sum is inherently progressive — evaluate the first two networks for a distant object, all of them up close. That’s a natural LOD scheme falling out of the representation rather than being authored.
Procedural modelling with a learned base. Take a scanned or learned surface as the coarse term and write your own analytic residual on top. Mixing learned and hand-written SDFs in one evaluation is a genuinely interesting compositional possibility.
The caveat: this is a graphics research paper, not a library. There’s no ComfyUI node and no npm package. The value right now is in understanding the technique — particularly the nested-band idea, which is transferable to hand-written multi-resolution SDF work whether or not you ever train an MLP.
Related Reading
- M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals — arXiv:2609.28684
- Inigo Quilez — distance functions reference
- Implicit Neural Representations with Periodic Activation Functions (SIREN) — Stanford
- NeuS / neural implicit surface reconstruction overview — arXiv
- arXiv cs.GR — Graphics listings