Reproduce “Inverse Design for Conditional Distribution Matching” on OpenResearch

About this paper

Ori Meidler, Shaul Tolkovsky, Or Zuk

The MLGD-F algorithm generalizes inverse design from point-wise objectives to the more challenging task of matching target output distributions using existing pretrained generative models. This inference-time approach, developed at the Hebrew University of Jerusalem, successfully generated outputs conforming to diverse target distributions across synthetic, MNIST, and Stable Diffusion benchmarks, achieving 22-33% lower MMD values compared to baselines in distribution matching scenarios.

Key insights

• Inverse design can be extended to distributional targets at inference time, leveraging existing pretrained generative models without requiring retraining.
• Practical implementation relies critically on fast, few-step conditional samplers to ensure computational tractability and memory efficiency for gradient computation.
• The method allows for controllable interpolation between sampling from the prior distribution and concentrating on inputs that minimize the distributional loss.

Summary by alphaXiv.