FerretNet: Lightweight Synthetic Detector

• FerretNet is a synthetic image detector that exploits local pixel dependencies using median-based reconstruction to identify subtle generative artifacts.
• It targets latent distribution deviations and decoding-induced smoothing effects through a compact 1.1M-parameter CNN, achieving 97.1% accuracy across 22 generator types.
• The method utilizes a 3x3 neighborhood for local reconstruction and dual-path grouped convolutions to efficiently simulate larger receptive fields and ensure robust forensic performance.

Searching arXiv for FerretNet and related synthetic image detection methods to ground the article with current papers. FerretNet is a synthetic image detection method centered on local pixel dependencies (LPD) and implemented as a lightweight neural architecture for open-world forensic generalization. It is introduced as a detector that targets two artifact types associated with modern image generators—latent distribution deviations and decoding-induced smoothing effects—and operationalizes these signals through local reconstruction based on neighboring pixels. In the reported formulation, FerretNet has only 1.1M parameters, is trained exclusively on the 4-class ProGAN dataset, and attains an average accuracy of 97.1% on an open-world benchmark comprising across 22 generative models, with the abstract reporting that it surpasses state-of-the-art methods by 10.6% (Liang et al., 25 Sep 2025).

1. Conceptual basis

FerretNet is motivated by the observation that the increasing realism of synthetic images generated by VAEs, GANs, and LDMs complicates synthetic image detection. The method isolates two classes of generation artifacts. The first is latent distribution deviations, described as a mismatch between the assumed and sampled latent distributions and associated with non-natural artifacts such as texture inconsistencies and structural detail loss. The second is decoding-induced smoothing effects, attributed to decoder choices such as large convolutional kernels, improper strides, and naive upsampling, which can introduce local texture loss, edge artifacts, over-smoothing, or aliasing (Liang et al., 25 Sep 2025).

The central modeling premise is that natural images exhibit strong local spatial regularities. FerretNet formalizes this through the Markov Random Field (MRF) assumption, under which the value of a pixel depends on its local neighborhood rather than on distant pixels:

$$P(x_{i, j}~|~\text{all}~x_{k, l\neq i, j}) = P(x_{i, j}~|~\{x_{k, l} : (k, l) \in \mathcal{N}_{i, j}\})$$

where $\mathcal{N}_{i, j}$ is an $n \times n$ window around $(i, j)$, excluding the center. Within this framing, real images are described as having strong, consistent local dependencies adhering to physical world statistics, whereas generative models can violate these dependencies in subtle ways, especially during decoding. A plausible implication is that FerretNet belongs to the class of forensic systems that prioritize micro-structural statistics over semantic content.

2. Local pixel dependency reconstruction

FerretNet’s feature extraction stage is based on local median-based reconstruction. For each pixel $x_{i,j}$ in an input image $I$, the method constructs a local patch $\mathcal{N}'_{i,j}$ of size $n \times n$ and applies zero-masking to the center pixel. It then computes the neighborhood median,

$y_{i, j} = \text{Median}(\mathcal{N}'_{i, j}),$

forming a reconstructed image $I'$ with entries $\mathcal{N}_{i, j}$0. The LPD map is defined as

$\mathcal{N}_{i, j}$1

This residual quantifies the deviation of each pixel from a local median estimate derived from its neighborhood (Liang et al., 25 Sep 2025).

The interpretation supplied for this construction is explicitly forensic. In natural images, the LPD map is described as tending toward low-amplitude, spatially smooth deviations. In synthetic images, edge or texture discontinuities can produce larger local deviations, which become visible in the LPD representation. FerretNet therefore converts detection into a residual-analysis problem: rather than classifying directly from RGB appearance, it classifies the discrepancy between observed pixels and locally reconstructed expectations.

Two implementation choices are emphasized. First, zero-masking the center pixel is said to reduce contamination from potentially anomalous synthetically generated values and to make the median more indicative of expected local appearance. Second, $\mathcal{N}_{i, j}$2 neighborhoods are reported to perform best empirically, balancing artifact sensitivity and noise rejection. Larger neighborhoods, specifically $\mathcal{N}_{i, j}$3 and $\mathcal{N}_{i, j}$4, are reported to degrade performance.

3. Network architecture

FerretNet is described as a lightweight neural network specialized for LPD maps rather than raw image semantics. Its topology begins with two $\mathcal{N}_{i, j}$5 convolution layers with batch normalization (BN) and ReLU. These are followed by four “Ferret Blocks.” Each block contains a primary path with a $\mathcal{N}_{i, j}$6 dilated grouped convolution using dilation rate 2 and group count equal to channel count, and a secondary path with a $\mathcal{N}_{i, j}$7 grouped convolution using the same grouping. The two outputs are concatenated and fused using $\mathcal{N}_{i, j}$8 convolution, BN, and ReLU, with residual connections used for stable training. The network ends with $\mathcal{N}_{i, j}$9 convolution, global average pooling, Dropout, and a fully connected layer for classification (Liang et al., 25 Sep 2025).

The architecture is explicitly optimized for parameter efficiency and throughput. The reported design leverages depthwise separable, grouped, and dilated convolutions to obtain an efficient receptive field expansion while keeping the parameter count at 1.1M. The dual-path arrangement is described as simulating a larger receptive field, “like a $n \times n$0 neighborhood,” but doing so efficiently.

The reported interpretive claim is that FerretNet avoids reliance on high-level semantic cues. In the description of the experiments, Grad-CAM is said to indicate focus on local anomalies rather than semantic concepts. This suggests that the architecture is tailored to capture spatially localized forensic evidence exposed by the LPD transformation.

4. Training regime and empirical evaluation

The training setup is deliberately constrained. FerretNet is trained on a 4-class ProGAN subset (ForenSynths) with 18K synthetic images/class for cars, cats, chairs, horses and 18K real images/class (from LSUN). Testing is conducted on several benchmarks: ForenSynths with 62K images spanning 8 GANs and multiple real-image sources; Diffusion-6-cls, covering 6 diffusion models with roughly 22K images; Synthetic-Pop, a diverse benchmark with 30K fake and 30K real images; and Synthetic-Aesthetic, comprising 40K CompVis/Stable diffusion fakes and 40K LAION-Aesthetics V2 real images. The evaluation metrics are Accuracy (ACC), Average Precision (AP), throughput (images/sec), and the protocol specifies that no pretraining is used (Liang et al., 25 Sep 2025).

The main comparative results reported in the data are as follows:

Test set	FerretNet (ACC / AP)	Comparison context
ForenSynths	95.9 / 99.3	Compared with FreqNet, NPR, FatFormer
Diffusion-6-cls	96.9 / 99.6	Compared with FreqNet, NPR, FatFormer
Synthetic-Pop	98.3 / 99.8	Compared with FreqNet, NPR, FatFormer
Synthetic-Aesthetic	97.3 / 99.6	Compared with FreqNet, NPR, FatFormer
Mean	97.1 / 99.6	Open-world average

The same report states that on Synthetic-Aesthetic, FerretNet reaches 772 images/sec, compared with 721 for NPR and 89 for FatFormer. Its parameter count is listed as 1.1M, whereas FatFormer is listed at 577M.

These numbers are significant because the detector is trained only on the small 4-class ProGAN dataset and is nevertheless reported to generalize across 22 generator types. The data also state that FerretNet achieves top performance on 22 generator types, including GAN and diffusion systems, which frames the method as an open-world detector rather than a closed-set classifier.

5. Ablations, robustness, and design justification

Several ablation results are presented to motivate the design. Increasing $n \times n$1 is not relevant here; that variable belongs to another method in the supplied material and does not apply to FerretNet. For FerretNet itself, the reported ablations concern the reconstruction operator and the network choice. Neighborhood size is one such axis: $n \times n$2 local median is reported to yield the best accuracy, while $n \times n$3 and $n \times n$4 reduce performance. The explanation given is that this is attributable to the small receptive fields in generator upsampling modules (Liang et al., 25 Sep 2025).

A second axis is the handling of the center pixel. Zero-masking is reported to consistently outperform exclusion or retention, both theoretically and empirically. A third axis is the summary statistic used in local reconstruction. Median-based extraction is reported to outperform max, min, avg, with the rationale that the median is more robust to outliers, including local synthetic artifacts.

A fourth axis is backbone selection. The report states that FerretNet with LPD outperforms both deeper Xception and ResNet50, which are said to have 20x more parameters, and also that LPD input boosts all backbones. Together with the Grad-CAM observations, these findings are used to argue that the principal source of discriminative power is the LPD representation and its alignment with local artifact detection rather than depth or semantic representation capacity.

6. Interpretation, scope, and common misconceptions

One possible misconception is that state-of-the-art synthetic image detection necessarily depends on large pretrained vision-language backbones or semantic priors. The reported protocol explicitly states no pretraining, and contrasts this with FatFormer, which is said to use CLIP. Another possible misconception is that open-world generalization requires training on a wide variety of generators. FerretNet is instead reported to be trained only on the 4-class ProGAN dataset while generalizing strongly to images from a wide variety of architectures (Liang et al., 25 Sep 2025).

A further misconception is that the detector works by learning generator-specific fingerprints in the narrow sense of model identity. The provided account does not frame the method that way. Rather, it attributes performance to a broader mechanism: local statistical analysis of deviations from natural-image neighborhood structure. This suggests that FerretNet’s effective domain is tied to artifact families that manifest as disruptions in texture continuity, edge coherence, and color transitions, irrespective of the particular generator family.

Within that scope, FerretNet is positioned as a practical detector for open-world deepfake and synthetic image forensics. The reported combination of 1.1M parameters, 772 images/sec on RTX4090, and strong performance on ForenSynths, Diffusion-6-cls, Synthetic-Pop, and Synthetic-Aesthetic indicates a design point that favors both deployment efficiency and cross-generator robustness. The evidence presented therefore supports a characterization of FerretNet as a detector grounded in MRF-inspired local dependency analysis and instantiated through a compact CNN specialized for LPD-space classification.