Why Are Conditional Generative Models Better Than Unconditional Ones?

TL;DR

This paper analyzes why conditional generative models outperform unconditional ones, revealing that proper data partitioning is key, and introduces self-conditioned diffusion models that improve performance without labels.

Contribution

The paper formally analyzes the advantage of conditional models and proposes self-conditioned diffusion models using clustering on features from self-supervised pretraining.

Findings

SCDM significantly improves unconditional model performance.

SCDM achieves a record-breaking FID of 3.94 on ImageNet 64x64.

SCDM outperforms or matches conditional models on benchmark datasets.

Abstract

Extensive empirical evidence demonstrates that conditional generative models are easier to train and perform better than unconditional ones by exploiting the labels of data. So do score-based diffusion models. In this paper, we analyze the phenomenon formally and identify that the key of conditional learning is to partition the data properly. Inspired by the analyses, we propose self-conditioned diffusion models (SCDM), which is trained conditioned on indices clustered by the k-means algorithm on the features extracted by a model pre-trained in a self-supervised manner. SCDM significantly improves the unconditional model across various datasets and achieves a record-breaking FID of 3.94 on ImageNet 64x64 without labels. Besides, SCDM achieves a slightly better FID than the corresponding conditional model on CIFAR10.