Research map / Image generation / Autoregressive and latent generation
What makes fake images detectable? Understanding properties that generalize
Lucy Chai, David Bau, Ser-Nam Lim, Phillip Isola
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 |
|---|---|---|---|
| use Xception Block 2 patch classifier instead of full Xception | fake image detector trained on 128px PGAN · PGAN 1024px | average precision: 97.28 -> 99.98 | Improved |
| use Resnet Layer 1 patch classifier instead of full Resnet-18 | detector on SGAN CelebA-HQ · SGAN CelebA-HQ | average precision: 64.80 -> 97.22 | Improved |
| shrink receptive field too far (Xception Block 2 to Block 1) (ablation) | patch classifier · PGAN 1024px | average precision: 99.98 -> 98.53 | Hurt |
| enlarge receptive field from Xception Block 2 to Block 5 (ablation) | patch classifier · PGAN 1024px | average precision: 99.98 -> 91.18 | Hurt |
| add parameters without enlarging receptive field (Extended Block 2) (ablation) | Xception Block 2 patch classifier · Glow, GMM, FFHQ SGAN | average precision: Glow 46.35 -> 39.23; GMM 91.38 -> 91.54; FFHQ SGAN 90.12 -> 87.82 | Mixed |
| add reprojected images to fake training set (ablation) | truncated patch classifiers · Glow | average precision: difference in AP of over 10% | Hurt |
| add random cropping during training (ablation) | Xception Block 1 patch classifier · Glow and FFHQ SGAN | average precision: Glow 82.39 -> 89.60; FFHQ SGAN 81.35 -> 79.00 | Mixed |
| save real and fake images with different formats (JPG vs PNG) (ablation) | Resnet-18 detector · SGAN CelebA-HQ | average precision: 64.80 -> 100.00 | Improved |
| use different interpolation for real and fake images (ablation) | Resnet-18 detector · Glow | average precision: 47.06 -> 99.76 | Improved |
| add blur and JPEG augmentation (p=0.5) | Resnet-18 detector · Glow | average precision: 47.06 -> 97.32 | Improved |
| adversarially finetune the PGAN generator against the classifier | first fake-image classifier · CelebA-HQ PGAN | accuracy: 100% -> below 65% | Hurt |
| use Resnet Layer 1 patch classifier instead of full Resnet-18 | detector trained on Face2Face · FaceForensics++ Deepfakes test | average precision: 55.43 -> 84.39 | Improved |
Closest papers by meaning
- Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.86)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.86)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.86)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now (CVPR 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.
