Jen-Kai (Kyle) Lin

Jen-Kai (Kyle) Lin

Research Assistant, AIIU Lab, Academia Sinica

Email CV GitHub LinkedIn

About

I work on structurally faithful 3D generation, with a focus on complete buildings: interior and exterior, multi-storey, enterable, and sheltering. Whether such a structure holds together is often invisible in images, so I measure it directly on the native 3D representation and use those measurements to evaluate, guide, and repair generation.

I am currently a research assistant at the AIIU Lab, Research Center for Information Technology Innovation, Academia Sinica, working with Prof. Jun-Cheng Chen.

News

Publications

A player-built cottage and a horizontal slice of its interior field
Shelter and Access: One-Sided, Falsifiable Interior Measures for Generated Voxel Buildings
Jen-Kai Lin. NeurIPS 2026 Workshop on Embodied Spatial Reasoning (ESR).
Paper Project page Code Dataset

Projects

A house built and repaired by an LLM
Measure What Was Built (ongoing)
An LLM builds Minecraft houses and repairs them from facts measured on the 3D data. Measured facts cut floating or mis-attached blocks from 102 to 26, while photos did not; a VLM pointed out only 7 of the 100 such blocks captured in tour photos.
Project page
Trunk cross-sections during sampling: without guidance the trunk stays disconnected, with CAST guidance it connects
CAST: Connectivity-Aware Sampling for Topology
Sampling-time guidance on a 3D voxel DDPM. A small connectivity scorer, applied in the middle of sampling, raises trunk connectivity from 6.5% to 30% without retraining the denoiser.
Report Code

Background