Attributing authorship to paintings is a historically complex task, and one of its main challenges is the limited availability of real artworks for training computational models. This study investigates whether synthetic images, generated through DreamBooth fine-tuning of Stable Diffusion, can improve the performance of classification models in this context. We propose a hybrid approach that combines real and synthetic data to enhance model accuracy and generalization across similar artistic styles. Experimental results show that adding synthetic images leads to higher ROC-AUC and accuracy compared to using only real paintings. By integrating generative and discriminative methods, this work contributes to the development of computer vision techniques for artwork authentication in data-scarce scenarios.
Figures & tables
Fig. 1: Overview of the image generation pipeline. Starting from Gaussian noise and a prompt (e.g., ”a full owhx painting”), the CLIP encoder conditions the U-Net denoising process in latent space. After multiple refinement steps, the latent image is decoded by a VAE into a full-resolution painting. This process is repeated to generate a synthetic dataset per artist.
Fig. 2: Patch sampling strategies. (A) Original image. (B) M1: baseline sampling with moderate overlap. (C) M2: denser sampling obtained by doubling the number of patches along both axes. Axis color reflects the level of overlap between samples.
Fig. 3: Examples of three synthetic images generated for each artist. The images reflect stylistic variations captured by the model, showcasing differences in composition, subject matter, and brushwork. Despite efforts to mitigate generation biases, some partial figures and cropped elements remain, likely influenced by the characteristics of the training dataset.
Fig. 4: Heatmap of ROC-AUC values for different training settings. The figure illustrates the variation in the model’s ability to discriminate between classes depending on the dataset used for training and testing. Higher values indicate better discrimination between authors.
Fig. 5: Heatmap of Accuracy values across training settings. The figure highlights differences in overall classification performance, showing how accuracy varies depending on whether the model was trained with real, synthetic, or mixed data.
Experiment
Metric
Most Accurate Artist
Best
Least Accurate Artist
Worst
Real-Only
ROC-AUC
TB
0.9593
JH
0.5976
Precision
GD
0.9644
JH
0.8346
Recall
GD
0.9919
JN
0.7163
F1-Score
GD
0.9780
JH
0.8288
Accuracy
GD
0.9573
JH
0.7193
Synthetic-Only
ROC-AUC
GM
0.9991
JN
0.9833
TABLE I: Performance comparison per experiment: best and worst artist-specific results across evaluation metrics
Synthetic image attribution (SIA) has become increasingly important with the rapid advancement of text-to-image generation models. However, accurately identifying the source model of a generated image remains challenging due to the growing similarity among modern diffusion-based generators and the presence of diverse post-processing operations. In this report, we present a multi-view and confusion-guided ensemble framework for the Synthetic Image Attribution Challenge of the DLMMDD Workshop at ICANN 2026. Our approach integrates multiple complementary architectures, including FFT-ConvNeXt, DINOv2, CLIP, and Xception, to capture diverse attribution cues from frequency, semantic, and forensic perspectives. To improve robustness against unknown degradations and image manipulations, extensive data augmentation strategies are employed during training, simulating realistic post-processing operations such as compression, resizing, grayscale conversion, and blur. Furthermore, we analyze the confusion patterns of the ensemble model and observe severe ambiguity between Stable Diffusion 3 and Stable Diffusion 3.5. To address this issue, we introduce a dedicated binary expert classifier that is selectively activated under low-confidence conditions. We additionally apply class-adaptive confidence calibration to improve the discrimination of challenging classes such as Tencent Hunyuan. The proposed framework achieved 99.53% on the public leaderboard and 99.20% on the private leaderboard. The source code and implementation details are publicly available at https://github.com/ZOMIN28/SIA.
Zuomin Qu
State Key Laboratory of HVDC, China Southern Power Grid Electric Power Research Institute, China
AI-generated image (AIGI) attribution presents a pressing challenge that goes beyond mere AIGI detection, aiming to identify the source model or technique responsible for a synthetic image. However, most previous source attribution methods operate in a closed-set manner, which necessitates retraining to recognize any novel category, preventing adaptation to the rapid evolution of image generation. In this work, we propose a new paradigm for synthetic image attribution, termed few-shot attribution. This paradigm targets the reliable identification of unseen generators using only limited samples, making it highly suitable for real-world applications. To facilitate this work, we construct OmniFake, a large-scale, well-categorized synthetic image dataset that contains 1.17 million images from 45 distinct generators. We further introduce OmniDFA (Omni Detector and Few-shot Attributor), a few-shot attribution baseline that not only assesses the authenticity of images but also determines their synthesis origins. Experiments demonstrate that OmniDFA exhibits excellent capability in few-shot attribution and achieves state-of-the-art generalization performance in AIGI detection. Our dataset and code are available at https://github.com/teheperinko541/OmniDFA.
Shiyu Wu, Shuyan Li, Jing Li +2
Institute of Automation, Chinese Academy of Sciences, Beijing, China · Beijing Academy of Artificial Intelligence, Beijing, China · University of Chinese Academy of Sciences, Beijing, China +3
The evolution and dissemination of AI-synthesized images is occurring at an unprecedented rate. Image generators are making rapid progress in their goal of perfectly imitating natural images, which also challenges image forensics. In this work, we exploit an underexplored cue in current generative models, namely their weakness to imitate color statistics of natural images. We first show that the LPIPS loss used for training image generators is less sensitive to chrominance than to luminance, which may lead to statistical discrepancies in the colors of synthetic images. Building on this observation, we then introduce six hand-crafted color transformations and a method to learn a task-optimized color transform to statistically expose generated images. These transformations can be used in various ways. First, we define color-sensitive features at pixel-level or patch-level. A simple, interpretable classifier achieves with these features an average generalization accuracy of 93.27% and strong robustness against six types of post-processing. Second, we demonstrate that the transformations exhibit characteristic visual noise patterns in natural and synthetic image areas, which enables an intuitive visual image evaluation. Third, we demonstrate that the transforms can enhance color patterns in generated images for improved multiclass attribution.