🤖 AI Summary
Existing image forgery detection methods suffer from poor generalization due to reliance on low-quality, narrow-domain training data. To address this, we introduce ImagiNet—the first high-resolution benchmark for synthetic image detection spanning diverse content types (photographs, paintings, faces, and miscellaneous objects), comprising 200K samples and supporting both binary authenticity classification and generative model attribution. We propose a multi-content balanced sampling paradigm and a dual-track evaluation framework. Our approach leverages ResNet-50 enhanced with SelfCon self-supervised contrastive learning, trained on synthetically generated images from both open-source and commercial diffusion models, paired with corresponding authentic images from public sources. The resulting model achieves AUC scores of 0.99 on both tasks, balanced accuracy of 86%–95%, strong robustness under JPEG compression and spatial rescaling, and zero-shot transfer performance that matches or exceeds state-of-the-art results on mainstream benchmarks.
📝 Abstract
Recent generative models produce images with a level of authenticity that makes them nearly indistinguishable from real photos and artwork. Potential harmful use cases of these models, necessitate the creation of robust synthetic image detectors. However, current datasets in the field contain generated images with questionable quality or have examples from one predominant content type which leads to poor generalizability of the underlying detectors. We find that the curation of a balanced amount of high-resolution generated images across various content types is crucial for the generalizability of detectors, and introduce ImagiNet, a dataset of 200K examples, spanning four categories: photos, paintings, faces, and miscellaneous. Synthetic images in ImagiNet are produced with both open-source and proprietary generators, whereas real counterparts for each content type are collected from public datasets. The structure of ImagiNet allows for a two-track evaluation system: i) classification as real or synthetic and ii) identification of the generative model. To establish a strong baseline, we train a ResNet-50 model using a self-supervised contrastive objective (SelfCon) for each track which achieves evaluation AUC of up to 0.99 and balanced accuracy ranging from 86% to 95%, even under conditions that involve compression and resizing. The provided model is generalizable enough to achieve zero-shot state-of-the-art performance on previous synthetic detection benchmarks. We provide ablations to demonstrate the importance of content types and publish code and data.