Detecting Diffusion-generated Images via Dynamic Assembly Forests

Mengxin Fu; Yuezun Li

cs.CVApr 10, 2026

Detecting Diffusion-generated Images via Dynamic Assembly Forests

Authors: Mengxin Fu, Yuezun Li

Organizations: School of Computer Science and Technology, Ocean University of China

Abstract

Diffusion models are known for generating high-quality images, causing serious security concerns. To combat this, most efforts rely on deep neural networks (e.g., CNNs and Transformers), while largely overlooking the potential of traditional machine learning models. In this paper, we freshly investigate such alternatives and proposes a novel Dynamic Assembly Forest model (DAF) to detect diffusion-generated images. Built upon the deep forest paradigm, DAF addresses the inherent limitations in feature learning and scalable training, making it an effective diffusion-generated image detector. Compared to existing DNN-based methods, DAF has significantly fewer parameters, much lower computational cost, and can be deployed without GPUs, while achieving competitive performance under standard evaluation protocols. These results highlight the strong potential of the proposed method as a practical substitute for heavyweight DNN models in resource-constrained scenarios. Our code and models are available at https://github.com/OUC-VAS/DAF.

Figures & tables

Figure 1: The overview of our proposed DAF model.

Figure 2: The detailed explanation of our proposed dynamic assembly strategy.

Figure 3: The detailed process of Task-specific Feature Extraction. (a) The input image is first divided into multiple patches. In each patch, we use HOG to extract spatial features. We also extract the frequency features of each patch and combine these two types of features together. (b) We employ sliding windows on batches to obtain multi-scale features.

Eval set

\downarrow

Train set

\to

LSUN-B ( ADM )

Imagenet ( ADM )

Detector

\to

CNNDet

AERO-

DIRE

Fake-

FIRE

DAF

CNNDet

AERO-

DIRE

Fake-

FIRE

DAF

BLADE

Inversion

(Ours)

BLADE

Inversion

(Ours)

Imagenet

ADM

73.4/57.6

99.1/98.3

99.4/96.4

100.0/99.8

100.0/100.0

96.5/99.9

99.6/99.3

100.0/100.0

99.5/100.0

SD‑v1

67.3/53.4

98.7/97.4

98.3/95.2

99.7/97.6

100.0/100.0

94.7/98.9

87.2/84.7

99.7/98.3

100.0/99.6

100.0/100.0

91.6/99.6

LSUN‑B

ADM

96.8/93.7

100.0/100.0

99.9/100.0

76.0/54.2

98.4/97.5

100.0/99.7

100.0/100.0

99.6/100.0

DDPM

75.3/55.1

98.9/97.8

100.0/100.0

100.0/99.8

99.9/100.0

68.4/45.1

99.1/98.2

100.0/97.7

100.0/100.0

99.7/100.0

Table 1: Detection performance of different methods (ACC%/AUC%).

Detector	Acc.(%)
CNNDet	60.11
FreDect	56.86
Fusing	57.07
GramNet	60.95
LNP	55.63
UniFD	55.62

Table 2: Detection performance of different methods on Chameleon dataset (ACC%).

Eval set $\downarrow$	Train set $\to$	LSUN-B ( ADM )		Imagenet ( ADM )
	Detector $\to$	Forensics	DAF	Forensics	DAF
	Detector $\to$	Forest	(Ours)	Forest	(Ours)
Imagenet	ADM	87.5/98.2	90.2/97.9	97.3/99.7	97.8/99.9
Imagenet	SD‑v1	84.5/91.6	89.6/94.0	86.3/96.2	79.7/98.3
LSUN‑B	ADM	92.0/99.0	98.0/100.0	94.0/98.1	97.5/99.9
	DDPM	91.0/98.6	98.9/99.9	94.4/98.3	97.2/99.9

Table 3: Comparison results of our method and ForensicsForest on a fair set (ACC%/AUC%).

Eval set $\downarrow$	Train set $\to$	LSUN-B ( ADM )			Imagenet ( ADM )
Eval set $\downarrow$	Detector $\to$	Single	Multi (Recalc.)	Multi (Avg.)	Single	Multi (Recalc.)	Multi (Avg.)
Imagenet	ADM	96.3/99.7	94.7/99.9	96.5/99.9	98.1/100.0	99.7/100.0	99.5/100.0
Imagenet	SD‑v1	93.7/98.6	94.3/98.6	94.7/98.9	90.1/98.5	84.7/99.3	91.6/99.6
LSUN‑B	ADM	99.8/99.9	99.9/100.0	99.9/100.0	99.5/99.9	98.3/100.0	99.6/100.0
	DDPM	99.7/99.9	99.8/100.0	99.9/100.0	99.5/99.9	98.9/100.0	99.7/100.0
	IDDPM	99.9/100.0	99.9/100.0	100.0/100.0	99.7/100.0	99.1/100.0	99.9/100.0

Table 4: Effect of multi-scale features.

Figure 4: Effect of the number of image patch partitions.

Figure 5: The effect of different

p

values on ACC (%) on ImageNet and LSUN-B datasets with three settings: Single, Multi (Recalc.), Multi (Avg.).

Figure 6: The visualization results trained on the LSUN-B dataset and tested on DALLE-2 and SD-v2 of the LSUN-B dataset.

Figure 7: The visualization results trained on the Imagenet dataset and tested on DALLE-2 and SD-v2 of the LSUN-B dataset.

Figure 8: Robustness under various perturbations, including Gaussian blur (top) and JPEG compression (bottom).

Eval set $\downarrow$	Train set $\to$	LSUN-B ( ADM )		Imagenet ( ADM )
Eval set $\downarrow$	Detector $\to$	$w.$	$wo.$	$w.$	$wo.$
Imagenet	ADM	94.2	96.5	99.5	99.5
Imagenet	SD‑v1	94.1	94.7	90.9	91.6
LSUN‑B	ADM	99.8	99.9	99.5	99.6
	DDPM	99.7	99.9	99.9	99.7
	IDDPM	99.8	100.0	100.0	99.9

Table 5: Discussion: ACC (%) performance of our method with(

w.

) or without(

wo.

) on-the-fly data augmentation.

Eval set $\downarrow$	Train set $\to$	LSUN-B ( ADM )		Imagenet ( ADM )
Eval set $\downarrow$	Detector $\to$	$w.$	$wo.$	$w.$	$wo.$
Imagenet	ADM	54.7	96.5	99.9	99.5
Imagenet	SD‑v1	40.3	94.7	98.3	91.6
LSUN‑B	ADM	100.0	99.9	50.0	99.6
	DDPM	99.9	99.9	56.6	99.7
	IDDPM	100.0	100.0	50.0	99.9

Table 6: Discussion: ACC (%) performance of our method with(

w.

) or without(

wo.

) ResNet18 feature extraction.

Explore similar work

FIND: A Simple yet Effective Baseline for Diffusion-Generated Image Detection
Mar 15, 2026Jie Li, Yingying Feng, Chi Xie +3Image Forgery Detection AI-Generated Image Detection
Xiamen University · Northeastern University · Tongji University +1
Fragile Reconstruction: Adversarial Vulnerability of Reconstruction-Based Detectors for Diffusion-Generated Images
Apr 14, 2026Haoyang Jiang, Mingyang Yi, Shaolei Zhang +4AI-Generated Image Detection Adversarial Robustness
Renmin University of China, Beijing, China · Tencent Inc., Shenzhen, China
FlowGuard: Towards Lightweight In-Generation Safety Detection for Diffusion Models via Linear Latent Decoding
Apr 9, 2026Jinghan Yang, Yihe Fan, Xudong Pan +1Diffusion Models Latent Diffusion Models
Fudan University, Shanghai, China · Fudan University, Shanghai Innovation Institute, Shanghai, China

Mar 15, 2026J/K move · Enter open · S save

FIND: A Simple yet Effective Baseline for Diffusion-Generated Image Detection

Jie Li, Yingying Feng, Chi Xie, Jie Hu+2

Xiamen University · Northeastern University · Tongji University+1

Image Forgery Detection AI-Generated Image Detection

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

FIND: A Simple yet Effective Baseline for Diffusion-Generated Image Detection

Fragile Reconstruction: Adversarial Vulnerability of Reconstruction-Based Detectors for Diffusion-Generated Images

FlowGuard: Towards Lightweight In-Generation Safety Detection for Diffusion Models via Linear Latent Decoding