Research map / Image generation / Autoregressive and latent generation

Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a low-level feature extractor on a pixel-difference pretext task (distinguishing originals from Stable Diffusion-denoised noisy versions) and detects AI-generated images as one-class anomalies under a GMM fit to real photos.

Problem

Detecting AI-generated images that generalize across rapidly evolving generators, including high-quality diffusion outputs, GANs and flow-based models, where binary classifiers trained on one generator fail to transfer.

Approach

Instead of learning generator artifacts, the paper uses a pretext task in which a feature extractor learns to tell original photos from versions noised and denoised by a pre-trained Stable Diffusion (used as a denoiser, not generator), then models only real-photo features with a Gaussian Mixture Model and flags low-likelihood images as AI-generated (one-class detection).

Key results

• On GenImage, the Ours-basic variant reaches 95.97% average accuracy versus 89.49% for DRCT, with consistent results across generators including BigGAN (95.71%).
• On DRCT-2M, Ours-ft achieves 90.86% average accuracy, while the best baseline DRCT scores 90.49% and Ours-basic scores 82.74%.
• Robustness: the method consistently outperforms DRCT under JPEG compression and downsampling on GenImage, though performance degrades as distortion increases.

Measured effects reported in the paper

Surprising findings

• Fine-tuning the feature extractor on GenImage did not improve performance there (95.27% vs 95.97% for the basic variant), which the authors attribute to accuracy saturation.
• Using a classic CNN denoiser (BF-CNN) or an early diffusion model (ADM) instead of Stable Diffusion for the pretext task yields much weaker detection (avg 78.36% and 84.61%).

Limitations

• Performance on DRCT-2M's SDv2-DR and SDXL-DR subsets (diffusion reconstructions from real images) remains weak for Ours-ft.
• Accuracy degrades under stronger JPEG compression and downsampling, although it still exceeds DRCT.
• Some generators in the zero-shot evaluation (Table 3) are still detected with low recall, e.g. StyleGAN2 at 83.53% for DRCT and 75.52% for the proposed method at the 75-percentile threshold.

Experimental setup

Datasets: YFCC100M GenImage DRCT-2M MSCOCO ImageNet

Benchmarks: GenImage DRCT-2M

Metrics: accuracy recall

Baselines: CNNSpot F3Net CLIP/RN50 GramNet De-fake Conv-B UnivFD DIRE LNP LGrad NPR DRCT

Builds on: Stable Diffusion DDIM DRCT UnivFD

Closest papers by meaning

• A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.88)
• Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
• Are Images Indistinguishable to Humans Also Indistinguishable to Classifiers? (CVPR 2025 · similarity 0.87)
• Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.86)
• Rethinking Peculiar Images by Diffusion Models: Revealing Local Minima’s Role (AAAI 2024 · similarity 0.86)
• Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.86)