Research map / Image generation / Autoregressive and latent generation
PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals
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 PiD, a generator-free AI-generated image detector that feeds a ResNet with pixelwise color-space quantization residuals (RGB→YUV→quantize→back to RGB, residual = input minus quantized output).
Problem
Detecting AI-generated images from unseen generators (GANs and diffusion models) is hard because detectors overfit to generator-specific artifacts, and reconstruction-based methods are computationally expensive and generator-biased.
Approach
Generative models optimize high-level semantics and overlook low-level residual components; PiD extracts a pixelwise decomposition residual by mapping RGB pixels to a color space (YUV), quantizing, and inverting, then uses the quantization residual as the detector input, with no generative model needed.
Key results
- PiD (YUV, round) reaches 96.22% mean accuracy on UniversalFakeDetect (ProGAN-trained), vs 95.89% for FatFormer and 93.21% for NPR.
- On GenImage (SDv1.4-trained), PiD achieves 98.0% mean accuracy, surpassing LARE2 by 11.8 points and C2P-CLIP by 2.2 points.
- On Self-Synthesis GAN dataset, PiD beats NPR by 1.2% mean accuracy and 2.8% mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use pixelwise YUV decomposition residuals instead of RGB input (ablation) | PiD detector · GenImage | accuracy: Base (RGB): 67.5, 98.7, 55.9, 61.6; YUV: 95.5, 99.6, 96.0, 96.2 | Helped |
| Use YUV rather than XYZ transformation (ablation) | PiD · GenImage | accuracy: Midjourney: 96.5 vs 95.5; SDv1.4: 97.8 vs 99.6; ADM: 96.5 vs 96.0; BigGAN: 81.6 vs 96.2 | Mixed |
| Use YUV rather than a diagonal transformation matrix (ablation) | PiD · GenImage | accuracy: Midjourney: 88.2 vs 95.5; SDv1.4: 94.2 vs 99.6; ADM: 88.3 vs 96.0; BigGAN: 88.4 vs 96.2 | Helped |
| Use YUV rather than a customized full-rank matrix (ablation) | PiD · GenImage | accuracy: Midjourney: 87.2 vs 95.5; SDv1.4: 96.5 vs 99.6; ADM: 89.4 vs 96.0; BigGAN: 91.7 vs 96.2 | Mixed |
| Scale the YUV transformation by 0.2 instead of 1.0 (ablation) | PiD · GenImage | accuracy: 0.2: 93.6, 99.2, 87.4, 91.2; 1.0: 95.5, 99.6, 96.0, 96.2 | Mixed |
| Scale the YUV transformation by 0.5 instead of 1.0 (ablation) | PiD · GenImage | accuracy: 0.5: 96.7, 98.7, 92.8, 97.9; 1.0: 95.5, 99.6, 96.0, 96.2 | Mixed |
| Scale the YUV transformation by 0.8 instead of 1.0 (ablation) | PiD · GenImage | accuracy: 0.8: 92.6, 98.8, 92.0, 91.3; 1.0: 95.5, 99.6, 96.0, 96.2 | Mixed |
| Scale the YUV transformation by 1.5 instead of 1.0 (ablation) | PiD · GenImage | accuracy: 1.5: 84.6, 93.1, 69.0, 79.6; 1.0: 95.5, 99.6, 96.0, 96.2 | Hurt |
| Scale the YUV transformation by 2.0 instead of 1.0 (ablation) | PiD · GenImage | accuracy: 2.0: 79.6, 93.3, 63.1, 76.4; 1.0: 95.5, 99.6, 96.0, 96.2 | Hurt |
| Use floor quantization instead of round quantization (ablation) | PiD · GenImage | accuracy: 98.0 vs 97.8 | Mixed |
| Use floor quantization instead of round quantization (ablation) | PiD · Self-Synthesis 9 GANs | accuracy: 94.7 -> 96.9 | Helped |
Surprising findings
- Even a simple diagonal scaling matrix (no color conversion) improves generalization over the RGB baseline on GenImage, per the ablation.
- Scaling the transform so residual intensity drops below 1 (scaling factor 1.5 or 2.0) sharply reduces generalization on GenImage.
Limitations
- Performance on some subsets varies (e.g., DALLE and GLIDE 100 steps lower accuracy, around 83-91%).
- Choice of transformation matrix matters; YUV performs best but no principled rule for selecting it is given.
- Compute and hardware details are not reported in the main text.
Experimental setup
Datasets: ForenSynths GenImage UniversalFakeDetect Self-Synthesis GAN dataset
Benchmarks: UniversalFakeDetect GenImage Self-Synthesis GAN dataset
Metrics: accuracy average precision
Baselines: CNN-Spot UnivFD LGrad FreqNet NPR FatFormer C2P-CLIP DIRE LARE2 DRCT Effort
Builds on: NPR DIRE FreqNet UnivFD
Closest papers by meaning
- STD-FD: Spatio-Temporal Distribution Fitting Deviation for AIGC Forgery Identification (ICML 2025 · similarity 0.86)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.86)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.85)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.85)
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.