Original title: Observation-Conditioned Latent Energy Priors for Sparse Implicit Neural Shape Completion
Paul Büschl, Ezequiel de la Rosa, Julia Wolleb, Julian McGinnis +2 · September 2026
Computers can learn to build complete 3D shapes from tiny bits of data. Often, an algorithm keeps its main shape decoder frozen and searches for a hidden digital code that matches a few sparse measurements. But when data points are too sparse, the algorithm can get confused. It finds a code that matches the sparse points yet creates bizarre, unrealistic shapes elsewhere. To fix this, researchers built an energy-based helper. This tool scores how well a code fits the actual observed points, acting alongside a standard baseline penalty. It guides the system toward realistic shapes without retraining the original decoder. When tested on sparse cell nuclei and medical shape data, this guided approach produced better, more realistic completions.
Imagine drawing a full picture from just a few scattered pencil marks. If the marks are in the top corner, you might sketch a bird, but if they're in the center, it could be a face. Without guidance, your guess might fit the marks but look nothing like a real scene. Here, the researchers give the system a rule that scores different drawings based on how well they match those specific pencil marks, ensuring the final picture is both complete and realistic according to the clues.
Doctors and engineers often work with incomplete 3D scans, like from MRI. This method helps reconstruct full shapes from fewer data points, leading to more accurate models without retraining the entire system.
The method is post-hoc and works only for frozen decoders, so it may not apply if the decoder is retrained or for different shape types. It is evaluated on specific cell-nucleus and MedShapeNet datasets, so performance on other data is not shown. The approach requires tuning a validation-selected L2 baseline, which might not generalize without adjustment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Observation-Conditioned Latent Energy Priors for Sparse Implicit Neural Shape Completion — Paul Büschl, Ezequiel de la Rosa, Julia Wolleb et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.