Research map / Image generation / Autoregressive and latent generation
Learning on Gradients: Generalized Artifacts Representation for GAN-Generated Images Detection
Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Yunchao Wei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| represent images with gradients from a pretrained StyleGAN-bedroom discriminator rather than raw images (ablation) | LGrad · Cross-model & category, 1-class (mean) | accuracy: 52.5 -> 86.3 | Improved |
| use gradients from a pretrained GAN discriminator | LGrad · Cross-model & category, 1-class (mean) | accuracy: 74.9 -> 86.3 | Improved |
| use a pretrained ProGAN discriminator rather than a randomly initialized one (ablation) | LGrad, CelebA-HQ to StyleGAN2 · CelebA-HQ to StyleGAN2 | accuracy: 52.4 -> 96.5 | Improved |
| use a GAN discriminator rather than a classification network as transformation model (ablation) | LGrad · Cross-model & category, 1-class (mean) | accuracy: 76.0 -> 86.3 | Improved |
| use InceptionV3 as the transformation model instead of StyleGAN-bedroom (ablation) | LGrad · Cross-model & category, 1-class (mean) | accuracy: 86.3 -> 56.9 | Hurt |
| increase training data from 4k to 36k images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category (mean) | accuracy: 75.9 -> 86.3 | Improved |
| increase training data from 36k to 72k images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category (mean) | accuracy: 86.3 -> 86.2 | No clear effect |
| blur test images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category | mean accuracy: 86.3 -> 78.5 | Hurt |
| crop test images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category | mean accuracy: 86.3 -> 85.0 | No clear effect |
| apply JPEG compression to test images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category | mean accuracy: 86.3 -> 69.5 | Hurt |
| add random noise to test images (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category | mean accuracy: 86.3 -> 69.1 | Hurt |
| apply combined blur, cropping, JPEG, and noise perturbations (ablation) | LGrad with StyleGAN-bedroom transformation · Cross-model & category | mean accuracy: 86.3 -> 75.9 | Hurt |
Closest papers by meaning
- Beyond Binary Classification: A Semi-supervised Approach to Generalized AI-generated Image Detection (AAAI 2026 · similarity 0.88)
- GenImage: A Million-Scale Benchmark for Detecting AI-Generated Image (NeurIPS 2023 · similarity 0.88)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.88)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.87)
- Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
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.
