Research map / Image generation / Autoregressive and latent generation

Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 5 citations (OpenAlex)

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 CoD, a lightweight (1.4M-parameter) AI-generated image detector that combines image features with a color-distribution feature obtained from color-depth quantization and noisy-averaging restoration, and introduces the FakeART cross-domain benchmark.

Problem

Detecting AI-generated images (GAN and diffusion) that generalize across generators and across domains (e.g., photos to artworks), where existing detectors degrade sharply.

Approach

Synthetic images show uneven color distributions; quantizing color depth, adding Gaussian noise, and averaging the quantized copies yields a restored image whose difference from the original reflects color uniformity. These difference maps are fed as a color feature alongside image features into a small ResNet classifier.

Key results

• On the GAN benchmark, CoD reaches 98.2% mean accuracy and 99.6% mAP, versus 96.0% / 98.7% for SAFE.
• On FakeART, CoD achieves 66.5% mean accuracy and 68.2% mAP, outperforming the next best (PatchCraft at 70.2 / 71.4 on some subsets; mean comparisons favor CoD over SAFE at 63.7 / 66.3 on the SD-1.5 column basis).
• The detector has 1.44M parameters; on GenImage it reaches 96.2% accuracy and 99.4% mAP.

Measured effects reported in the paper

Surprising findings

• Color features alone achieve 89.1% on GenImage and 59.3% on FakeART, and pretrained CLIP features alone lower FakeART accuracy to 54.8%.
• Adding CLIP features on top of image and color features reduces FakeART accuracy (64.3 vs 66.5), suggesting pretrained features hurt cross-domain generalization.

Limitations

• Color quantization and noise averaging add processing time as the number of noisy images grows.
• FakeART accuracy (around 66.5%) remains modest, showing cross-domain detection is still hard.
• Theoretical analysis relies on approximating the Fourier series by a single sine term.

Experimental setup

Datasets: ProGAN-LSUN training set COCO LSUN WikiArt Forensic Synthetics (Wang et al.) Ojha benchmark GenImage

Benchmarks: Forensic Synthetics (GAN) Ojha diffusion benchmark GenImage FakeART

Metrics: accuracy mean average precision

Baselines: CNNDet FreDect LGrad UFD PatchCraft FreqNet NPR FatFormer SAFE

Builds on: CLIP UniFD FatFormer NPR SAFE

Closest papers by meaning

• Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · similarity 0.86)
• Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.85)
• CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
• FerretNet: Efficient Synthetic Image Detection via Local Pixel Dependencies (NeurIPS 2025 · similarity 0.85)
• Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.85)
• Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.85)