Research map / Image generation / Autoregressive and latent generation
A Closer Look at Fourier Spectrum Discrepancies for CNN-Generated Images Detection
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace zero-insert scaling in the final generator step with nearest interpolation (ablation) | DCGAN, LSGAN, and WGAN-GP · CelebA 128x128 | successful detection rate: 84.82±3.72% -> 0.09±0.03%; 88.16±3.98% -> 0.34±0.08%; 99.75±0.14% -> 0.14±0.05% | Helped |
| Replace zero-insert scaling in the final generator step with bilinear interpolation (ablation) | DCGAN, LSGAN, and WGAN-GP · CelebA 128x128 | successful detection rate: 84.82±3.72% -> 0±0%; 88.16±3.98% -> 0.1±0%; 99.75±0.14% -> 0.2±0.12% | Helped |
| Use nearest or bilinear rather than zero-insert scaling in the final step (ablation) | GAN generators · CelebA 128x128 | high-frequency spectral consistency | Helped |
| Use zero-insert scaling in the final step (ablation) | GAN generators · CelebA 128x128 | high-frequency spectral consistency | Hurt |
| Vary the final convolution kernel from 5x5 to 3x3 (ablation) | Nearest- and bilinear-scaling GANs · CelebA 128x128 | high-frequency spectral consistency | Mixed |
| Vary the final convolution kernel from 5x5 to 7x7 (ablation) | Nearest- and bilinear-scaling GANs · CelebA 128x128 | high-frequency spectral consistency | No clear effect |
| Increase the number of final convolution blocks from one to three (ablation) | GAN generators · CelebA 128x128 | high-frequency spectral consistency | Mixed |
| Replace zero-insert scaling with nearest interpolation (ablation) | DCGAN, LSGAN, and WGAN-GP · LSUN Bedrooms 128x128 | successful detection rate: 97.88±1.04% -> 1.5±2.09%; 95.79±2.07% -> 0.62±0.43%; 99.87±0.13% -> 0.03±0.05% | Helped |
| Use bilinear interpolation in StarGAN's last scaling step (ablation) | StarGAN · CelebA 256x256 | successful detection rate: 64.3±2.7% -> 0±0% | Helped |
| Use nearest interpolation in StarGAN's last scaling step (ablation) | StarGAN · CelebA 256x256 | high-frequency spectral consistency | Hurt |
| Use nearest interpolation rather than the baseline final transpose convolution (ablation) | DCGAN, LSGAN, and WGAN-GP · CelebA 128x128 | FID: 88.6 -> 87.52; 73.26 -> 70.69; 60.6 -> 48.69 | Mixed |
| Use bilinear interpolation rather than the baseline final transpose convolution (ablation) | DCGAN, LSGAN, and WGAN-GP · CelebA 128x128 | FID: 88.6 -> 84.65; 73.26 -> 78.66; 60.6 -> 52.18 | Mixed |
Closest papers by meaning
- Fourier Spectrum Discrepancies in Deep Network Generated Images (NeurIPS 2020 · similarity 0.91)
- Any-Resolution AI-Generated Image Detection by Spectral Learning (CVPR 2025 · similarity 0.87)
- Spectrum Translation for Refinement of Image Generation (STIG) Based on Contrastive Learning and Spectral Filter Profile (AAAI 2024 · similarity 0.86)
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.86)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain (ICML 2025 · 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.