Research map / Image generation / Autoregressive and latent generation
FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies
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 FerretNet, a 1.1M-parameter detector that feeds zero-masked neighborhood-median residuals (local pixel dependency maps) into a lightweight depthwise/dilated CNN, trained only on 4-class ProGAN, and reports 97.1% ACC across 22 generators.
Problem
Synthetic image detection that generalizes to unseen GAN, VAE, and latent diffusion generators (including high-fidelity modern models) while remaining parameter- and compute-efficient.
Approach
Represent images by local pixel dependencies: the residual between each pixel and a zero-masked median of its neighborhood (motivated by Markov Random Field assumptions), which exposes texture and edge inconsistencies left by decoders; a tiny CNN with dilated grouped convolutions classifies these residuals.
Key results
- FerretNet reports 97.1% ACC / 99.6% AP averaged across four test sets with 1.06M parameters and 772.1 FPS on an RTX 4090, versus FatFormer's 86.1% ACC (a 492.59M-parameter model) on the same aggregate.
- On ForenSynths FerretNet gets 95.9% mean ACC, below FatFormer's 98.4%; on Diffusion-6-cls it gets 96.9% ACC, above FatFormer's 95.0%.
- Replacing raw input with LPD (3x3 neighborhood) raises mean ACC from 86.9% to 97.1% for FerretNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 3×3 LPD input instead of raw image input (ablation) | FerretNet · Four test sets, mean | ACC: 86.9 -> 97.1 | Helped |
| Increase the LPD neighborhood from 3×3 to 5×5 (ablation) | FerretNet · Four test sets, mean | ACC: 97.1 -> 93.9 | Hurt |
| Increase the LPD neighborhood from 3×3 to 7×7 (ablation) | FerretNet · Four test sets, mean | ACC: 97.1 -> 82.8 | Hurt |
| Zero-mask the center pixel rather than exclude it (ablation) | FerretNet LPD extraction · Four test sets, mean; 3×3 neighborhood | ACC: 95.3 -> 97.1 | Helped |
| Zero-mask the center pixel rather than retain it (ablation) | FerretNet LPD extraction · Four test sets, mean; 3×3 neighborhood | ACC: 93.3 -> 97.1 | Helped |
| Use neighborhood median instead of maximum (ablation) | FerretNet LPD extraction · Four test sets, mean; 3×3 neighborhood | ACC: 93.6 -> 97.1 | Helped |
| Use neighborhood median instead of average (ablation) | FerretNet LPD extraction · Four test sets, mean; 3×3 neighborhood | ACC: 92.2 -> 97.1 | Helped |
| Use neighborhood median instead of minimum (ablation) | FerretNet LPD extraction · Four test sets, mean; 3×3 neighborhood | ACC: 91.8 -> 97.1 | Helped |
| Replace raw image input with LPD input (ablation) | ResNet50 · Four test sets, mean | ACC: 75.0 -> 81.1 | Helped |
| Replace raw image input with LPD input (ablation) | Xception · Four test sets, mean | ACC: 89.8 -> 95.1 | Helped |
| Scale FerretNet down to 0.13 M parameters (ablation) | FerretNet-B · Four test sets, mean | ACC: 97.1 -> 93.1 | Hurt |
| Scale FerretNet up to 21.51 M parameters (ablation) | FerretNet-B · Four test sets, mean | ACC: 97.1 -> 96.6 | Hurt |
Surprising findings
- Scaling FerretNet up to 21.5M parameters slightly lowers mean accuracy (96.6% vs 97.1%), which the authors attribute to overfitting.
- Under JPEG compression (Q=75) FerretNet drops to 50.2% ACC on ForenSynths, near chance, and FatFormer also drops sharply.
Limitations
- Robustness to compression-altered images is not well explored; JPEG compression at Q=75 drops ACC to about 50%.
- False positives on highly structured real textures (e.g., rock formations) and false negatives on smooth, homogeneous backgrounds.
- Results on the Synthetic-Pop and Synthetic-Aesthetic benchmarks are reported by the authors' own dataset construction, which limits independent comparison.
Experimental setup
Datasets: ForenSynths ProGAN 4-class training set (LSUN real images) ForenSynths test set Diffusion-6-cls Synthetic-Pop Synthetic-Aesthetic (SAC + LAION-Aesthetics V2)
Benchmarks: ForenSynths Diffusion-6-cls Synthetic-Pop Synthetic-Aesthetic
Metrics: accuracy average precision parameters FLOPs FPS
Baselines: Wang et al. (CNNSpot) F3Net FrePGAN BiHPF LGrad Ojha et al. FreqNet NPR FatFormer SAFE CO-SPY
Builds on: NPR FatFormer CLIP Markov Random Fields
Closest papers by meaning
- Layer Consistency Matters: Elegant Latent Transition Discrepancy for Generalizable Synthetic Image Detection (CVPR 2026 · similarity 0.89)
- RealNet: Efficient and Unsupervised Detection of AI-Generated Images via Real-Only Representation Learning (AAAI 2026 · similarity 0.87)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.86)
- Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis (CVPR 2025 · similarity 0.85)
- Enhancing Consistency-Based Image Generation via Adversarialy-Trained Classification and Energy-Based Discrimination (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.