Robustness of Synthetic Image Detec-tors in the Wild: Cross-Domain Evaluation and Visual Literacy-Based Error Analysis

Synthetic Image Detection Across Different Generative Models

This paper investigates synthetic digital image detection across different generative models. The task is relevant due to rapid development of generative artificial intelligence, which enables creation of visual content that is difficult to distinguish from real imagery and increases risks related to misinformation, visual forgery, and declining trust in digital media. The study compares several detection approaches: CLIP-based semantic representations, FFT-based frequency-domain features, fusion models combining semantic and frequencydomain information, and ensembles of detectors. Experiments were conducted on tinyGenImage, a subset of GenImage, using BigGAN and Stable Diffusion v1.5 images for training and validation. Performance across different sources was evaluated on images generated by Midjourney, Wukong, and GLIDE, which were not used during training. Results show that CLIP-based models provide a strong baseline, while FFT-only models perform weaker as standalone detectors. Fusion models did not consistently improve over CLIP baselines, whereas ensembling achieved the best overall performance and improved classification quality across different generative sources.

Visual Distortion Detection in UGC Images Using Large Multimodal Models

The localized depiction of perceptual quality has long been a crucial, yet underexplored, challenge in image quality assessment (IQA). Existing approaches based on large multimodal models (LMMs) predominantly rely on text-driven supervised fine-tuning (SFT). However, this training paradigm exhibits notable limitations in detection accuracy. Moreover, synthetically distorted images, which are often used as the primary training data source, show a significant generalization gap when deployed in real-world scenarios; thus, the \textbf{synthetic-to-authentic (\textit{S2A})} problem represents a critical challenge. Motivated by these issues, we propose \textbf{\textit{VIGIL}}, which leverages the LMM architecture for precise visual distortion detection. From a candidate pool of over 1000K samples, we construct the \textbf{\textit{VIGIL-140K}} training set, which consists of over 140K distorted images. These images are obtained through rigorous quality filtering and carefully crafted distortion injection, covering 8 major synthetic distortion categories. Our model leverages different layers of the large language model (LLM) decoder, treating them as \textit{multiple detectors} that perform synchronous distortion detection using multi-level features. Additionally, we retain distortion cues from predictions assigned to the non-distortion class, which helps mitigate the ambiguous foreground-background (\textit{FG-BG}) separation commonly encountered in the \textit{S2A} problem. After post-processing, our model consistently outperforms strong baselines on both in-domain synthetic distortion detection and \textit{S2A} tasks.

Generated Images Are Easier to Forget: A Machine Unlearning Perspective for Synthetic Image Detection

Robust detection of generated images is critical to counter the misuse of generative models. Existing methods primarily depend on learning from human-annotated training datasets, limiting their generalization to unseen distributions. In contrast, large-scale vision models (LVMs) pre-trained on web-scale datasets exhibit exceptional generalization power through exposure to diverse distributions, offering a transformative paradigm for this task. However, our experimental results reveal that LVMs pre-trained on natural-image-dominated data can effectively capture the features of both natural and generated images, yielding comparably low losses and thus limited discriminative capacity between them. This prompts a key question: When and how do LVMs exhibit different behaviors when capturing features of natural and generated images? This investigation reveals an insight: during unlearning, LVMs exhibit disparate forgetting dynamics with feature degradation for generated images escalating faster than natural ones. Inspired by the disparate dynamics, we introduce two detection methods: 1) data-free detection, which prunes model parameters to induce unlearning without data access, and 2) data-driven detection, which optimizes LVMs to unlearn knowledge tied to generated images. Extensive experiments conducted on various benchmarks demonstrate that our unlearning-based approach outperforms conventional detection methods. By recasting the detection task as a problem of machine unlearning, our work establishes a new paradigm for generated image detection.

Benchmarking and Improving Perceptual Straightening for Robust AI-Generated Video Detection

The proliferation of text-to-video generative models—including commercial systems such as Sora, Veo, Runway Gen-3, and Kling—makes automated detection of AI-generated videos an urgent problem in multimedia forensics. We present a study with two interleaved contributions: a cross-family benchmark and two improvements to the detection pipeline. For the benchmark, we evaluate three methodologically distinct detector families on three public datasets (VidProM, DeepAction, and DeCoF_T2V), spanning diverse open-source and commercial generators: ReStraV, a geometry-supervised detector grounded in the perceptual straightening hypothesis; D3, a training-free detector based on second-order XCLIP temporal statistics; and DeMamba, a supervised Bidirectional Mamba module on frozen XCLIP features. ReStraV and DeMamba achieve broadly comparable global performance on large datasets (VidProM AUC 0.985/0.989; DeCoF_T2V AUC 0.990/0.992), while D3 remains weak as a stand-alone detector under a fixed detection threshold calibrated on VidProM and transferred unchanged to the other benchmarks (AUC 0.534/0.410/0.454 on VidProM/DeepAction/DeCoF_T2V). Per-generator analysis reveals complementary failure modes: DeMamba recovers several hard commercial generators where ReStraV struggles, whereas ReStraV remains competitive in low-data settings at substantially lower computational cost. To bridge these weaknesses, we propose two improvements. First, Rich384 enriches ReStraV’s compact geometric descriptor with DINOv2 temporal embeddings, strengthening ranking quality on large benchmarks and recovering generators that geometry alone misses (AUC 0.995 on VidProM and 0.998 on DeCoF_T2V, with DeepAction AUC decreasing to 0.805). Second, we propose GurAI, our transparent logistic late-fusion method that combines Rich384 and DeMamba logits, raising threshold-free AUC to 0.998/0.999 on VidProM and DeCoF_T2V while preserving interpretability. On the small DeepAction cohort, fusion improves selected per-generator fake recalls at the cost of elevated REAL false positives (AUC 0.769; REAL accuracy 0.460).Our analysis suggests that transparent late fusion can exploit complementary detector strengths more effectively than architectural redesign alone when facing generator diversity.

Evaluating domain generalization of product image quality assessment models using AI-generated images

AI-generated Images Challenge Visual Trust in High-risk Scenarios

Rapid advances in image generation are eroding the evidentiary value of visual content in settings where authenticity can affect public safety and personal reputation. Yet existing detection benchmarks rarely examine synthetic images in public- and individual-safety contexts, where misleading visual content may carry substantial risks. Here we introduce SafeIMG, a safety-oriented benchmark spanning 12 public- and individual-safety scenarios generated using GPT Image 2. Unlike benchmarks centred on generic imagery and image-level labels, SafeIMG evaluates not only whether detectors recognise synthetic images, but also whether their decisions reflect human-identified anomalies. To this end, SafeIMG provides human annotations that localise suspicious regions and explain local artefacts and higher-level commonsense or physical inconsistencies. We evaluate specialized synthetic-image detectors and vision-language models (VLMs), and find that neither provides reliable detection. The strongest VLM identifies only 49.5% of generated images, whereas the best specialised detector identifies 33.1%, compared with 81.7% accuracy for human evaluators. Model explanations cover only 29.8\% of human-annotated anomalies and predominantly capture local defects in text, faces and hands. Their coverage falls to 15.0% for commonsense conflicts and 12.0% for physical inconsistencies, while detection performance deteriorates further after dissemination-induced image degradation. These findings show that current detectors lack the accuracy, explanatory alignment and robustness needed to evaluate AI-generated images reliably across public- and individual-safety settings.