Synthesize Privacy-Preserving High-Resolution Images via Private Textual Intermediaries

We propose SPTI (Private Textual Intermediaries), a training-free framework for generating high-resolution differentially private (DP) synthetic images. Instead of operating directly in the image domain, SPTI converts private images into textual descriptions, applies DP text generation via a modified Private Evolution algorithm, and reconstructs images using text-to-image models — all with off-the-shelf models and no fine-tuning. SPTI significantly outperforms prior DP image synthesis methods, achieving FID = 26.71 on LSUN Bedroom and FID = 33.27 on MM-CelebA-HQ at ε=1.0.

Code: https://github.com/MarkGodrick/SPTI

Recommended citation: Wang, H., Lin, Z., Yu, D., & Zhang, H. (2025). "Synthesize Privacy-Preserving High-Resolution Images via Private Textual Intermediaries." Advances in Neural Information Processing Systems (NeurIPS). 2025.
Download Paper