HFI: A unified framework for training-free detection and implicit watermarking of latent diffusion model generated images

TL;DR

The paper tackles the challenge of training-free detection of AI-generated images from latent diffusion models (LDMs), addressing generalization across diverse generators. It introduces High-frequency Influence (HFI), a score that leverages the autoencoder of an LDM as a downsampling/upsampling kernel and quantifies aliasing in reconstructions via a low-pass filter, with an ensemble approach over multiple autoencoders. Empirically, HFI outperforms existing training-free detectors on challenging benchmarks (GenImage, SynthBuster, DiffusionFace) and remains competitive with training-based methods, while offering substantial test-time efficiency. Additionally, HFI can serve as an implicit watermarking mechanism for tracing images produced by a specified LDM, delivering near-perfect attribution with large speedups over optimization-based baselines. The work advances practical, scalable detection and ownership tracing for LDM-generated content in real-world, data-free settings.

Abstract

Dramatic advances in the quality of the latent diffusion models (LDMs) also led to the malicious use of AI-generated images. While current AI-generated image detection methods assume the availability of real/AI-generated images for training, this is practically limited given the vast expressibility of LDMs. This motivates the training-free detection setup where no related data are available in advance. The existing LDM-generated image detection method assumes that images generated by LDM are easier to reconstruct using an autoencoder than real images. However, we observe that this reconstruction distance is overfitted to background information, leading the current method to underperform in detecting images with simple backgrounds. To address this, we propose a novel method called HFI. Specifically, by viewing the autoencoder of LDM as a downsampling-upsampling kernel, HFI measures the extent of aliasing, a distortion of high-frequency information that appears in the reconstructed image. HFI is training-free, efficient, and consistently outperforms other training-free methods in detecting challenging images generated by various generative models. We also show that HFI can successfully detect the images generated from the specified LDM as a means of implicit watermarking. HFI outperforms the best baseline method while achieving magnitudes of

Figures (5)

• Figure 1: Problem setup of $\text{HFI}$.(Left) Setup of training-based AI-generated image detection methods. Such methods train and test on the same real data distribution. Furthermore, the framework can be costly when detecting images produced by large-scale text-to-image generative models. (Right) Pipeline of our proposed $\text{HFI}$. $\text{HFI}$ operates only on the test time and can be computed efficiently via the autoencoder of the LDM.
• Figure 2: Motivation of $\text{HFI}$.(a) Sampled data from the ImageNet Deng09ImageNet dataset. (b) Reconstruction through the autoencoder of the Stable Diffusion Rombach22SD v1.1 model. We can observe obvious distortions in the high-frequency details. (c) Histogram of AEROBLADE ricker24aeroblade experimented in toy dataset. (d) Histogram of $\text{HFI}$ experimented in toy dataset.
• Figure 3: Samples of toy dataset.(a) Real ImageNet Deng09ImageNet data. (b),(c),(d) AI-generated data. SDv1.5 (b), SDv2-base (c)Rombach22SD, and Kandinsky (d)Razzhigaev23Kandinsky are applied for generation, respectively.
• Figure 4: Visualization of the edge-cases.(a) Visualization of the ImageNet data where AEROBLADE outputs the smallest uncertainty. (b) Visualization of the SDv1.4-generated data where AEROBLADE outputs the highest uncertainty. We mark the sample where $\text{HFI}$ also fails.
• Figure 5: Performance of $\text{HFI}$, AEROBLADE, and B-HFI under corruption.