Source-linked AI summary
A Multi-View and Confusion-Guided Ensemble Framework for Robust Synthetic Image Attribution
Zuomin Qu
TL;DR
Synthetic image attribution remains challenging because similar diffusion generators and post-processing can obscure source-specific evidence. The paper addresses this with a multi-view ensemble, augmentation, confusion-guided expert refinement, and calibration, achieving 99.53% on the public leaderboard and 99.20% on the private leaderboard.
Problem
Synthetic image attribution is difficult because modern diffusion generators have overlapping characteristics and post-processing can weaken forensic traces.
Method
The framework combines FFT-ConvNeXt, DINOv2, CLIP, and Xception with augmentation, a selectively activated SD3/SD3.5 binary expert, and class-adaptive calibration.
Results
99.53% on the public leaderboard and 99.20% on the private leaderboard were achieved by the final solution.
Takeaways & Limitations
The framework addresses inter-class ambiguities, particularly between highly similar generators, while releasing code and training protocols for reproducibility.
Abstract
from arXiv · showhide
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.
1 Introduction
Synthetic image attribution is increasingly important as realistic generated images become harder to distinguish from real content. This framework combines complementary models, augmentation, confusion-guided refinement, and calibration to address overlapping generator fingerprints and post-processing effects.
- Synthetic image attribution identifies the source generative model of increasingly realistic synthetic images and supports digital media forensics and AI security.
- Robust attribution is difficult because diffusion generators share visual characteristics while post-processing can weaken source-specific forensic traces.
- The framework combines FFT-ConvNeXt, DINOv2, CLIP, and Xception with extensive augmentation to capture complementary cues and improve robustness to unknown test-time processing.Augmentations simulate operations such as compression, resizing, grayscale conversion, and enhancement.
- Confusion analysis motivates a binary expert classifier for Stable Diffusion 3 versus Stable Diffusion 3.5 and class-adaptive calibration for difficult classes such as Tencent Hunyuan.
2 Methodology
The framework combines complementary frequency, semantic, and forensic views with augmentation, weighted logit ensembling, and targeted refinement for ambiguous source classes.
- FFT-ConvNeXt Architecture: FFT-ConvNeXt concatenates RGB and filtered frequency representations to jointly exploit spatial structures and generator-specific forensic traces.A high-pass mask emphasizes high-frequency traces, while the RGB and frequency branches are combined before classification.
- Multi-View Ensemble Strategy: Four complementary models capture frequency, semantic, and forensic attribution cues through FFT-ConvNeXt, DINOv2, CLIP, and Xception.FFT-ConvNeXt captures frequency artifacts, DINOv2 global structure, CLIP semantic representations, and Xception forensic patterns.
- Multi-View Ensemble Strategy: Weighted logit fusion combines the models’ predictions, while K-fold averaging improves inference stability and robustness.Each model contributes weighted logits, and predictions from different folds are averaged during inference.
- Robust Augmentation: Extensive augmentation simulates resizing, cropping, brightness and contrast changes, grayscale conversion, JPEG compression, and hidden test-time degradations.The pipeline is designed to address unknown post-processing operations and distribution shifts, with JPEG simulation and grayscale conversion reported as especially useful.
- Confusion-Guided Expert Refinement: Validation analysis identifies severe SD3–SD3.5 confusion, prompting selective binary-expert refinement under a confidence threshold.The framework also applies class-adaptive confidence calibration to Tencent Hunyuan and uses pretrained weights with partial fine-tuning.
3 Experimental Setup
The experiments use a balanced 10-model synthetic-face dataset with undisclosed post-processing designed to weaken source traces, and evaluate complementary ensemble models under these conditions.
- Dataset: 10 open-source text-to-image models contribute 1,000 synthetic face images each to the balanced 10,000-image dataset.The training set contains 7,000 labeled images and the test set contains 3,000 unlabeled images.
- Dataset: Undisclosed compression, resizing, cropping, blur, grayscale, and super-resolution operations weaken generation traces and increase inter-class ambiguity.
- Training Configuration: The experiments use batch size 16, up to 50 epochs, early stopping with 10 patience epochs, and backbone-specific input resolutions.FFT-ConvNeXt uses 256 × 256 inputs, DINOv2 and CLIP use 224 × 224, and Xception uses 288 × 288.
- Training Cost: The four ensemble models have different computational profiles, with FFT-ConvNeXt incurring the highest complexity because of frequency-domain processing and feature fusion.Xception has lower FLOPs but relatively high training time because of implementation overhead and high-resolution forensic extraction.
4 Results
The weighted logits ensemble combines complementary models and achieves the strongest reported leaderboard performance, while targeted refinement improves predictions for ambiguous source classes.
- Inference Configuration: The weighted logits ensemble assigns weights [0.2, 0.4, 0.3, 0.1] to FFT-ConvNeXt, DINOv2, CLIP, and Xception, respectively.DINOv2 and CLIP receive higher weights for generalization under distribution shifts and post-processing perturbations.
- Public Leaderboard Performance: DINOv2 achieves the best standalone performance among single-model approaches, while FFT-ConvNeXt provides strong frequency-domain attribution cues.
- Public Leaderboard Performance: 0.995333 on the public leaderboard and 99.20% on the hidden private leaderboard are achieved by the proposed ensemble system.The private score indicates performance beyond the public evaluation subset.
- SD3/SD3.5 Expert Classifier: The SD3/SD3.5 expert classifier raises the public leaderboard score from 0.992666 to 0.995333 under the confusion-guided refinement ablation.The expert is selectively activated for the highly confused Stable Diffusion 3 and Stable Diffusion 3.5 classes.
- Tencent Hunyuan Calibration: Tencent Hunyuan confidence calibration raises the public leaderboard score from 0.993333 to 0.995333 for a partially ambiguous class.
5 Reproducibility Statement
The report provides a stepwise reproduction workflow covering baseline training, expert training, ensemble-logit generation, and final prediction refinement.
- Reproduction Workflow: The workflow first trains four baseline classifiers with K-fold cross-validation and stores their weights in checkpoints/.
- Reproduction Workflow: The Confusion-Guided Expert classifier is trained separately, with its weights also stored in checkpoints/.
- Reproduction Workflow: K-fold ensemble logits are generated on the test set and saved in the logits result/ directory.
- Reproduction Workflow: The final step performs weighted ensembling, activates the expert when applicable, calibrates Tencent Hunyuan confidence, and saves predictions in results/.
- Reproducibility: The released code, configuration files, and preprocessing routines are intended to ensure full reproducibility on the challenge dataset.
6 Conclusion
The report concludes that combining complementary logits-level models with targeted confusion-guided refinement addresses ambiguities in synthetic image attribution, with code and protocols released for reproducibility.
- Conclusion: The framework combines FFT-ConvNeXt, DINOv2, CLIP, and Xception at the logits level with targeted refinement for highly similar generators.