Compositional Energy-Based Inference-Time Scaling for Multi-Scale Microstructure Generation

Chengxi Zhang, Yu Yao, Zhenting Qi, Yilun Du, Ju Li

Keywords: Generative Modeling, Energy-Based Models, Diffusion Models, Inference-Time Scaling, Scientific Machine Learning

TL;DR: CEBS composes energy functions at inference time to decompose conditions and match target distributions, improving multi-scale scientific image generation.

Abstract: Generating scale-consistent scientific images requires satisfying physical constraints over combinatorial conditions, while standard point-wise selection can collapse the joint distribution. We introduce Compositional Energy-Based Scaling (CEBS), an inference-time scaling framework based on compositional energy minimization. CEBS decomposes a jointly trained conditional diffusion model into condition-specific energy gradients and recombines them for controllable out-of-distribution generation without retraining. It further defines a set-level energy that penalizes mean shift and covariance mismatch, selecting candidate subsets that preserve distributional statistics rather than optimizing samples independently. CEBS is general to conditional generators, requires no auxiliary discriminator, and improves scientific image generation under interpretable statistical and physical constraints.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 242