AIGI-Align

To address this problem, this paper proposes AIGI-Align, an instance-level semantic alignment dataset and benchmark for AIGI detection. To mitigate semantic bias from the data perspective, AIGI-Align is constructed under an instance-level semantic alignment constraint. The core idea is to use each real image as a semantic reference and generate corresponding fake images that are aligned with it in fine-grained semantic dimensions.

Real images are selected from ImageNet and filtered according to image resolution to ensure basic visual quality. A multimodal large model is used to verify whether the image content matches the target category, ensuring semantic category correctness and reducing category mismatch. Unlike conventional datasets that rely on coarse category labels such as "a photo of [class]" as generation prompts, this work introduces a vision-language model to generate fine-grained structured descriptions for each real image.

These descriptions cover four visual dimensions: subject description, photographic style, image details, and background elements. These image-level descriptions are then used as text prompts to drive generative models, so that the generated images are constrained by the semantic content of the corresponding real images rather than by a coarse class label alone. Based on these prompts, 12 generative models are used to synthesize fake images, covering U-Net-based latent diffusion models, diffusion Transformers, rectified-flow Transformers, cascaded diffusion models, and a text-to-video model used in a single-frame setting.

After generation, the fake images are checked by a multimodal large model for semantic category consistency, and samples with obvious visual artifacts are removed through manual inspection. The dataset contains 532 000 high-resolution images, including 70 000 real images and 462 000 generated images. It covers 14 semantic categories and 12 generative models, providing both instance-level semantic alignment training data and a controlled benchmark for evaluating cross-generator generalization.

Experiments show that AIGI-Align improves the degree of semantic alignment between real and fake images, alleviates the interference of semantic bias in the decision process of detectors, and improves their cross-generator generalization ability. Semantic bias analysis further indicates that existing detectors may use semantic differences between real and fake images for shortcut classification. We conduct a systematic evaluation on eight representative AIGI detectors, including CNN-based methods, frequency- or transformation-based methods, and methods based on pretrained visual representations.

Evaluation is performed on ForenSynths, UniversalFakeDetect, GenImage, and the proposed AIGI-Align benchmark. Experimental results show that, under the conventional training setting, some detectors exhibit a performance drop in the semantic alignment evaluation scenario. This suggests that their original performance may be partially supported by semantic shortcuts rather than by robust recognition of forgery traces.

When real and fake images are more consistent in semantic content, the semantic differences available to detectors are weakened, making the evaluation more focused on whether detectors can capture generation-related traces. After introducing instance-level semantic alignment training data, most detectors achieve improved cross-generator detection performance on both AIGI-Align and public benchmarks. These results indicate that instance-level semantic alignment data can reduce detectors' dependence on semantic content differences and encourage them to learn more general forgery-related cues.

This paper presents AIGI-Align, a large-scale instance-level semantic alignment dataset for AIGI detection. In terms of training, instance-level semantic alignment data can guide detectors to focus more on generation-related forgery traces and reduce their reliance on semantic content differences during real-fake discrimination, thereby improving their generalization ability to unseen generative models. In terms of evaluation, the AIGI-Align test set provides a controlled scenario in which semantic variables are constrained, making it possible to evaluate the actual generalization performance of detectors when semantic shortcuts are weakened.