Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation

2026-05-30

AI Summary

This work investigates the effective integration of self-supervised representation learning with one-step generative modeling to enhance both generation quality and matching stability. The proposed method generates samples in the reconstruction space while aligning distributions in a frozen self-supervised semantic feature space using Sinkhorn divergence as an approximation of the Wasserstein distance. A key contribution is the demonstration that the choice of semantic feature space critically influences the tractability of distribution matching, leading to the introduction of matching stability as a novel criterion for feature selection. Notably, the study shows that the features used during training and evaluation need not be identical. Experiments on ImageNet demonstrate up to a 39-fold reduction in FID, confirming that specific families of self-supervised features substantially improve both generative performance and stability.

Abstract

Generative modeling and self-supervised representation learning (SSL) optimize structurally different objectives: generative training rewards distributional fidelity, while SSL rewards semantic coherence. Yet recent work repeatedly finds that SSL features improve generative training, though the mechanism of this synergy remains unclear. Here, we study the benefits of SSL in generative modeling in the framework of one-step generation where the role of representation is explicit: frozen SSL features are used to match generated samples to real data. We use the Sinkhorn divergence in that feature space, providing a tractable surrogate for the Wasserstein distance, the population-level discrepancy approximated by Fréchet-style evaluation metrics (such as FID). We find that this objective becomes highly effective when computed in a semantically structured SSL feature space (a 39$\times$ reduction in ImageNet FID). We trace this behavior primarily to matching estimation: semantic SSL features that suppress nuisance reconstruction details induce a more compact geometry, making distribution matching more tractable. As a consequence, the best training SSL features need not match the features used by the evaluation metric. In particular, we show that using Inception as the feature extractor can improve FID while degrading matching stability and sample quality, revealing a form of metric hacking. Using extensive experiments on ImageNet, we identify which SSL feature families lead to best generation performance and show that matching stability is a quantitative criterion for selecting them. Code is available at https://github.com/Genentech/semantic-transport-generation.

Problem

• generative modeling
• self-supervised learning
• distribution matching
• semantic representation
• evaluation metrics

Innovation

• one-step generation
• self-supervised learning
• Sinkhorn divergence
• semantic feature space
• distribution matching