New framework accurately attributes synthetic images, distinguishing between Stable Diffusion versions

Researchers have developed a novel framework for attributing synthetic images, achieving high accuracy on a challenge dataset. Their approach combines multiple AI architectures, including FFT-ConvNeXt, DINOv2, CLIP, and Xception, to analyze images from various perspectives like frequency, semantics, and forensics. To enhance robustness against image manipulations, they incorporated extensive data augmentation simulating real-world post-processing. The framework specifically addresses confusion between Stable Diffusion 3 and Stable Diffusion 3.5 by employing a dedicated binary classifier and class-adaptive confidence calibration, ultimately scoring 99.20% on the private leaderboard. AI

IMPACT This research advances the field of synthetic image attribution, crucial for detecting AI-generated content and ensuring authenticity in digital media.

RANK_REASON The cluster describes a research paper detailing a new framework for synthetic image attribution, including model architectures and performance metrics. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• DINOv2
• DLMMDD Workshop
• FFT-ConvNeXt
• ICANN 2026
• Stable Diffusion 3
• Stable Diffusion 3.5
• Tencent Hunyuan
• Xception

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →