[Paper Review] FIND: A Simple yet Effective Baseline for Diffusion-Generated Image Detection

Jie Li, Yingying Feng+4 more|arXiv (Cornell University)|Mar 15, 2026

Advanced Neuroimaging Techniques and Applications0 citations

TL;DR

FIND trains a binary classifier by adding Gaussian noise to real images and labeling noisy versions as synthetic, removing the need for reconstruction models and achieving state-of-the-art accuracy and speed on GenImage.

ABSTRACT

The remarkable realism of images generated by diffusion models poses critical detection challenges. Current methods utilize reconstruction error as a discriminative feature, exploiting the observation that real images exhibit higher reconstruction errors when processed through diffusion models. However, these approaches require costly reconstruction computations and depend on specific diffusion models, making their performance highly model-dependent. We identify a fundamental difference: real images are more difficult to fit with Gaussian distributions compared to synthetic ones. In this paper, we propose Forgery Identification via Noise Disturbance (FIND), a novel method that requires only a simple binary classifier. It eliminates reconstruction by directly targeting the core distributional difference between real and synthetic images. Our key operation is to add Gaussian noise to real images during training and label these noisy versions as synthetic. This step allows the classifier to focus on the statistical patterns that distinguish real from synthetic images. We theoretically prove that the noise-augmented real images resemble diffusion-generated images in their ease of Gaussian fitting. Furthermore, simply by adding noise, they still retain visual similarity to the original images, highlighting the most discriminative distribution-related features. The proposed FIND improves performance by 11.7% on the GenImage benchmark while running 126x faster than existing methods. By removing the need for auxiliary diffusion models and reconstruction, it offers a practical, efficient, and generalizable way to detect diffusion-generated content.

Motivation & Objective

• Motivate robust detection of diffusion-generated images and address generalization and efficiency gaps in reconstruction-based detectors.
• Propose a reconstruction-free baseline by exploiting distributional differences between real and synthetic images through noise augmentation.
• Theoretically justify why noise-perturbed real images resemble diffusion-generated data in Gaussian fitting.
• Empirically demonstrate strong accuracy and extreme speed advantages on GenImage across multiple generators.

Proposed method

• Hypothesize real images are harder to fit with Gaussian distributions than synthetic ones in the context of diffusion reconstruction.
• Add Gaussian noise to real images during training and label these noisy samples as synthetic to induce a distribution-focused discriminative feature.
• Train a simple binary classifier on real, noisy-real, and synthetic images without any reconstruction or auxiliary diffusion models.
• Prove that noise Augmentation reduces Fisher information, making real-noise distributions resemble synthetic distributions for the classifier.
• Evaluate FIND on GenImage with ViT-B/16 and RN50 backbones, comparing against reconstruction-based baselines.
• Demonstrate that inference requires no reconstruction and is significantly faster than prior methods.

Figure 1: Comparison between FIND and previous noise-based methods. (a) DIRE utilizes complete reconstruction by adding noise and then denoising it with a Diffusion Model (DM) in multiple steps; (b) LaRE 2 compresses the noise addition and denoising steps into a single step within the latent space;

Experimental results

Research questions

• RQ1 Can a reconstruction-free binary classifier effectively distinguish real vs. diffusion-generated images?
• RQ2 Does Gaussian noise augmentation of real images shift their distribution to resemble diffusion-generated data in a way that improves detection?
• RQ3 How does FIND generalize across multiple diffusion generators compared to reconstruction-based detectors?

Key findings

• FIND achieves higher average accuracy than prior methods on GenImage with both RN50 and ViT-B/16 backbones.
• With ViT-B/16, FIND attains an average accuracy of 88.35% across generators, outperforming FatFormer and LaRE 2.
• FIND eliminates the need for reconstruction models and diffusion-based reconstruction, yielding a 126× faster inference compared to state-of-the-art methods.
• Training with random Gaussian noise on real images substantially improves performance, especially when combining noise on real images with noise on both real and synthetic data.
• Optimal noise magnitude (epsilon) was found at 50, balancing image integrity and discriminative strength.
• FIND demonstrates stable, cross-generator performance, showing strong generalization across eight diffusion generators in GenImage.

Figure 2: The training framework of FIND. Gaussian noise is added to real images, and these perturbed versions are labeled as synthetic within the training process. This enables FIND to mitigate the reconstruction model dependency, and learn the core distributional differences between real and synth