Title: LAMP: Data-Efficient Linear Affine Weight-Space Models for Parameter-Controlled 3D Shape Generation and Extrapolation

Abstract:

Creating high-fidelity 3D geometries under explicit parameter constraints is a cornerstone of engineering design. However, existing approaches frequently demand extensive datasets and struggle to offer dependable control when operating outside the training distribution. To address these limitations, we present LAMP, a framework designed for controllable and interpretable 3D generation that prioritizes data efficiency. LAMP operates by aligning signed distance function (SDF) decoders; this is achieved by overfitting each individual example from a common initialization. Subsequently, the model creates new designs by resolving a parameter-constrained affine mixing problem within this aligned weight space.

To enhance the reliability of the generation process, we introduce a linearity-mismatch safety metric. This tool identifies instances where mixed decoders deviate from the valid local regime. We assessed LAMP’s capabilities using DrivAerNet++ and BlendedNet, as well as several industry-level vehicle categories, such as convertibles, SUVs, and sports cars. The framework facilitates controlled interpolation with datasets as small as 50 samples, allows for safe extrapolation reaching 100% beyond the training boundaries, and supports performance-guided optimization under fixed parameters. In terms of extrapolation accuracy, data efficiency, and adherence to parameter fidelity, LAMP surpasses baselines including conditional autoencoders and Deep Network Interpolation (DNI). These findings highlight LAMP’s potential to advance safe, data-efficient, and controllable 3D generation, offering significant benefits for dataset creation, design exploration, and performance-driven optimization.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC