Research map / Image generation / Autoregressive and latent generation

Any-Resolution AI-Generated Image Detection by Spectral Learning

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 23 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.

Proposes SPAI, a self-supervised spectral AI-generated image detector that learns the real-image frequency distribution via masked frequency reconstruction and flags deviations with spectral reconstruction similarity and spectral context attention for any-resolution inputs.

Problem

Detecting AI-generated images that generalize to generators unseen during training, at native resolutions, while remaining robust to online perturbations.

Approach

Model the spectral distribution of real images with a self-supervised masked frequency reconstruction ViT (MFM), treat generated images as out-of-distribution via spectral reconstruction similarity between original and low/high-pass reconstructions, and fuse patch-level spectral vectors with spectral context attention so any-resolution images can be processed without resizing.

Key results

• SPAI achieves 91.0 average AUC across 13 generators, a 5.5 point absolute improvement over the second-best method (RINE, 85.5).
• Replacing the MFM spectral backbone with CLIP ViT-B/16 or DINOv2 yields 87.6 and 87.5 AUC versus 91.0 for SPAI.
• Removing spectral pretraining drops AUC to 52.5, and removing SRS drops it to 71.0.

Measured effects reported in the paper

Surprising findings

• Without spectral pretraining of the backbone, performance collapses to 52.5 AUC, near chance.

Limitations

• Fails to detect derivative AI-generated images (screenshots, photos of screens, prints) where spectral artifacts are distorted.
• Detection relies on image signal that is degraded by compression and noisy channels.

Experimental setup

Datasets: ImageNet Corvi et al. 180k real and latent diffusion images RAISE FODB COCO Open Images GigaGAN images Glide SD1.3 SD1.4 DALLE2 SD2

Benchmarks: 13 generative models across 5 real-image sources (RAISE, FODB, ImageNet, COCO, Open Images)

Metrics: AUC

Baselines: NPR Dire CNNDetect FreqDetect Fusing LGrad UnivFD GramNet DeFake PatchCraft DMID RINE

Builds on: Masked Frequency Modeling (MFM) DMID UnivFD Vision Transformer

Closest papers by meaning

• Training-free Detection of AI-generated images via Cropping Robustness (NeurIPS 2025 · similarity 0.87)
• A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
• A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.85)
• Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.84)
• CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.84)
• Acoustic-to-Hyper-Spectral: Hyper-Spectral Image Construction from Frequency Spectrums Through Simulated Annealing (Student Abstract) (AAAI 2025 · similarity 0.84)