Event

ReGDiff uses controlled latent space guided diffusion to generate metamaterial voxel structures

1 reports · 1 sources

The team of Dawei Zhou at Virginia Tech, together with Meta researchers, proposed ReGDiff in a NeurIPS 2026 paper for generating metamaterial voxel structures that conform to geometric constraints. The method has two steps: first, an RAS mechanism is used to organize the autoencoder latent space into a form in which good and bad structures are clearly separated and good structures are evenly spread out; then, short-range repulsion (SRR) pushes latent variables away from the neighborhoods of known samples during sampling, and it takes effect only at inference time. The team selected 10,000 samples from a 3D graph dataset to construct the MetaTruss dataset and uniformly scaled MetaShell to 48³. Compared with the strongest baseline, Yang et al., geometric plausibility improved by an average of 8.9% on the two datasets, novelty increased by 46.4%, and diversity increased by 128.6%. When generating 1,000 structures, ReGDiff took 31 seconds, close to the 28 seconds of Yang et al., while DiT-3D and 3D-CDM required 326 seconds and 813 seconds, respectively.