Can Self-Supervised Probabilistic Representations Improve Image Reconstruction?

What image information remains recoverable from a probabilistic model pretrained through non-contrastive self-supervision without pixel reconstruction? Our experiments show that decoding its frozen representations recovers visual information, although pixel fidelity remains below that of conventional variational autoencoders (VAEs). We subsequently compare fine-tuning a VAE encoder initialized from this self-supervised model with training a separate VAE whose prior is the same model's frozen image-conditioned distribution over latent codes (posterior distillation). Fine-tuning improves pixel fidelity over frozen decoding and conventional VAEs but reduces class separability relative to the frozen self-supervised model. Across CIFAR-10, CIFAR-100, and Tiny ImageNet, posterior distillation improves perceptual similarity, raises peak signal-to-noise ratio by 0.31–0.66 dB, and increases linear classification accuracy over conventional VAEs. Cross-dataset experiments show that the reconstruction benefit extends beyond the pretraining dataset.