FusionDetect: Hybrid AI Image Detector

• FusionDetect is a feature-fusion detector that integrates complementary features from CLIP and DINOv2 to detect synthetic images under both unseen generators and semantic domains.
• It fuses high-level semantic and fine-grained structural cues by concatenating frozen model features and training a lightweight 4-layer MLP, achieving up to 97.38% accuracy on diverse benchmarks.
• The method introduces a two-axis generalization framework for evaluating fake image detectors, emphasizing robustness against both cross-generator and cross-domain shifts.

Searching arXiv for FusionDetect and closely related papers to ground the article. FusionDetect is a feature-fusion detector for AI-generated image detection introduced in "Redefining Generalization in Visual Domains: A Two-Axis Framework for Fake Image Detection with FusionDetect" (Amanzadi et al., 7 Oct 2025). It is defined by two linked contributions: a detector that combines two frozen vision foundation models, CLIP and DINOv2, and a reformulation of synthetic-image detection as a two-axis generalization problem over both generators and semantic domains. Within that framing, FusionDetect is designed to operate when both the train-time generators and the train-time semantic domains are disjoint from those encountered at test time, while remaining robust under common image perturbations (Amanzadi et al., 7 Oct 2025).

1. Problem formulation and scope

A central claim of the work is that most prior evaluation of fake-image detectors has focused on cross-generator generalization, but that this viewpoint is too limited because synthetic-image detection also requires cross-semantic, or cross-domain, generalization (Amanzadi et al., 7 Oct 2025). The paper formalizes this as a two-axis problem: if $G_{train}$ and $S_{train}$ denote the train-time generators and semantic domains, and $G_{test}$ and $S_{test}$ denote the test-time sets, the target regime is one in which $G_{train} \cap G_{test} = \emptyset$ and $S_{train} \cap S_{test} = \emptyset$. In this formulation, a detector is not regarded as universal merely because it transfers across unseen generators; it must also remain effective on unseen content domains.

The empirical motivation for that reframing is supplied through t-SNE projections showing that GenImage, ImagiNet, and Chameleon form non-overlapping clusters in feature space. The reported implication is that a detector trained on one dataset may fail on others because of domain gap, even when the underlying generators are similar or the same (Amanzadi et al., 7 Oct 2025). This directly challenges the common simplification that cross-generator transfer alone is an adequate proxy for real-world generalization.

2. Architecture and feature-fusion mechanism

FusionDetect is a feature-fusion model built on two frozen vision foundation models with complementary roles. CLIP is used for semantics, encoding high-level semantic, contextual, and object-level information, and is described as strong for cross-semantic generalization. DINOv2 is used for structure and texture, capturing fine-grained structure, textures, and artifacts, and is described as assisting cross-generator generalization (Amanzadi et al., 7 Oct 2025).

Given an image $I$, the encoders produce feature vectors $E_{CLIP}(I) \in \mathbb{R}^{d_{clip}}$ and $E_{DINO}(I) \in \mathbb{R}^{d_{dino}}$. Fusion is implemented by concatenation,

$z_f = [E_{CLIP}(I) \ \| \ E_{DINO}(I)] ,$

followed by a lightweight multi-layer perceptron, which is the only trainable component of the network. The default setting is a 4-layer MLP. Training uses standard binary cross-entropy,

$S_{train}$0

with $S_{train}$1 (Amanzadi et al., 7 Oct 2025).

The encoders are kept frozen during training to preserve their robust, general-purpose features and to prevent overfitting. The stated intuition is that the CLIP branch contributes semantic breadth, while the DINOv2 branch contributes sensitivity to subtle statistical cues left by generators; the fused representation is intended to address both axes of generalization simultaneously (Amanzadi et al., 7 Oct 2025).

3. Training protocol and benchmark construction

The training set contains 60,000 images, consisting of 30k real and 30k fake, sampled for diversity from GenImage, ImagiNet, and Chameleon. For cross-generator evaluation, only SD1.4 and SD2.1 are used in training. Optimization is performed for 10 epochs with AdamW on a single RTX 3090. Because the backbones are frozen, the setup is described as enabling fast experimentation (Amanzadi et al., 7 Oct 2025).

To improve robustness, 10% of training images undergo random JPEG compression and Gaussian blur. Evaluation is reported with accuracy, average precision, and the mean and standard deviation of metrics across datasets, where lower standard deviation is interpreted as robustness across the two axes. Full per-dataset and per-class accuracies are reported, and robustness stress testing is performed with JPEG and blur perturbations (Amanzadi et al., 7 Oct 2025).

The paper also introduces OmniGen, a new benchmark intended to reflect both semantic and generator diversity. OmniGen incorporates 12 state-of-the-art generators, including GPT-4o, Midjourney v7, FLUX 1, SD3.5, Dreamshaper, and community finetunes. The synthetic portion contains 11,550 images spanning broad styles and content, matched to 1,000 real images from Unsplash (Amanzadi et al., 7 Oct 2025). Established benchmarks used alongside OmniGen are GenImage, ImagiNet, and Chameleon.

4. Empirical performance

On the established benchmarks GenImage, ImagiNet, and Chameleon, FusionDetect reports the best mean accuracy, 80.86%, and the best average precision, 87.40%, across all datasets. It also reports the lowest standard deviation, which is presented as evidence of consistency across datasets. On ImagiNet and Chameleon, performance is reported as approximately 10% better than other top detectors, while the abstract summarizes the overall gain as 3.87% more accurate than the closest competitor and 6.13% more precise on average on established benchmarks (Amanzadi et al., 7 Oct 2025).

On OmniGen, FusionDetect reports a top accuracy of 97.38% and the lowest standard deviation, 3.26%. The abstract further states that it achieves a 4.48% increase in accuracy on OmniGen. The benchmark-level interpretation given in the paper is that the detector remains consistent across all 12 state-of-the-art generators, including unseen, closed-source, and community generators (Amanzadi et al., 7 Oct 2025).

These results are presented as support for the underlying two-axis formulation. The paper’s argument is not only that FusionDetect performs well on average, but that its variance across datasets and generators is unusually low relative to prior methods, which is essential if the detector is to be used under domain shift rather than only under generator shift (Amanzadi et al., 7 Oct 2025).

5. Robustness, ablations, and comparative interpretation

Under JPEG compression and blur, artifact-based detectors such as SSP, NPR, and DNF are reported to degrade drastically. FusionDetect, by contrast, is reported to maintain stable performance, with accuracy staying above 80% at strong JPEG compression and blur. The paper interprets this as evidence that the detector relies on more robust high-level features rather than brittle artifact cues alone (Amanzadi et al., 7 Oct 2025).

Ablation studies indicate that the fusion of CLIP and DINOv2 features outperforms either feature source alone. Classifier complexity from 1 to 5 layers has little effect, and the reported conclusion is that a 4-layer classifier is sufficient and that the feature backbone is the key determinant rather than head depth. Feature upscaling with FeatUp did not improve results, and the paper states that the raw encoder features are already optimal in this setting (Amanzadi et al., 7 Oct 2025).

The comparative framing is explicit. Artifact-based approaches are characterized as brittle and poor in cross-generator and cross-domain generalization, especially under perturbation. Prior foundation-model approaches are characterized as better on semantic generalization, but still suboptimal when used alone. FusionDetect’s position is therefore a hybrid one: it is neither a hand-crafted artifact detector nor a single-backbone foundation-model detector, but a frozen-feature fusion model intended to combine semantic and structural evidence in a single representation (Amanzadi et al., 7 Oct 2025).

6. Relation to broader fusion-detection research and implications

Within the broader literature, the phrase “fusion detection” refers to several distinct technical paradigms. "E2E-MFD: Towards End-to-End Synchronous Multimodal Fusion Detection" addresses visible-infrared image fusion and object detection through synchronous joint optimization (Zhang et al., 2024). "Features Fusion for Dual-View Mammography Mass Detection" uses deformable-attention feature fusion and object-level correspondence in dual-view mammography (Varlamova et al., 2024). "DBF: Dynamic Belief Fusion for Combining Multiple Object Detectors" and the earlier "Dynamic Belief Fusion for Object Detection" are score-level late-fusion methods based on Dempster-Shafer theory (Lee et al., 2022, Lee et al., 2015). FusionDetect differs from these by operating in fake-image detection and by defining fusion as the concatenation of frozen foundation-model features rather than multimodal pixel fusion or belief-level detector fusion.

The paper’s recommendations follow directly from this distinction. It argues that evaluation for universal AI image detection should be performed along both axes—cross-generator and cross-domain—using benchmarks with broad semantic and generative diversity. It further recommends frozen foundation-model features as a practical design choice for efficiency and generalization, with only a minimal classifier trained on top (Amanzadi et al., 7 Oct 2025).

In that sense, FusionDetect is not only a detector but also part of a proposed evaluation framework. The work introduces OmniGen as a benchmark, advocates two-axis generalization as the relevant criterion, and presents the detector as a concrete instantiation of that framework. The code and dataset are reported as available at http://github.com/amir-aman/FusionDetect (Amanzadi et al., 7 Oct 2025).