Time Step Generating: A Universal Synthesized Deepfake Image Detector

Abstract

The rise of high-fidelity text-to-image diffusion models has made synthetic images increasingly indistinguishable from real ones, posing serious threats in digital security and media integrity. Existing detection methods often rely on reconstruction-based pipelines, which are computationally expensive and brittle on out-of-distribution data. We propose Time Step Generating (TSG), a universal synthetic image detector that leverages a pre-trained diffusion model as a feature extractor. By inputting images at a fixed diffusion timestep, TSG captures semantic and structural differences in noise prediction behavior between real and generated images — all within a single forward pass, enabling lightweight and effective classification. To eliminate the reliance on the manually chosen timestep hyperparameter, we further introduce TSG++, an enhanced version that consolidates multi-timestep diffusion features through lightweight fine-tuning. TSG++ learns to align features across all timesteps, producing a unified representation that improves both robustness and generalization without additional inference cost. Experiments on GenImage and challenging multimedia datasets demonstrate that TSG and TSG++ outperform prior methods in both accuracy and efficiency, offering a strong and adaptable solution for diffusion-based synthetic image detection.

ZIYUE ZENG

Ph.D. Student at Waseda University

Pursuing socially meaningful, grounded research with enduring passion and diligence.