Research map / Image generation / Text-to-image diffusion
Are High-Quality AI-Generated Images More Difficult for Models to Detect?
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.
Shows that AI-generated images with higher human-preference scores are easier for existing detectors to catch, explains this via prompt length and low-level image traits, and improves patch-based detectors by picking the most detectable patch.
Problem
Whether high-quality AI-generated images preferred by humans are harder for AIGI detectors to identify, given that existing benchmarks contain low-quality, randomly generated images while real-world synthetic images tend to be high quality.
Approach
Uses pretrained human preference models (ImageReward, HPS v2) as quality scorers on a newly built high-quality, diverse dataset and finds a positive correlation between quality and detector accuracy; then links this to prompt length and low-level image features via clustering and linear regression, and applies the regression to select the most detectable patch for existing detectors.
Key results
- Images with higher ImageReward/HPS v2 scores are consistently easier to detect across four generators, with cluster-level Pearson correlation between quality and accuracy positive but not strong.
- Selecting the highest-scored ('Easiest') patch improved average accuracy of end-to-end detectors, e.g., RINE from 75.0 to 74.2 (ACC) and SuSy from 76.0 to 77.8, while DRCT (Conv-B) rose from 75.4 to 76.4.
- For fingerprint-based detectors, the 'Hardest' patch strategy was best, e.g., NPR ACC rose from 71.6 (center) to 76.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select the highest-scoring patch using the accuracy regression model instead of a random patch | End-to-end AIGI detectors · GenImage, DRCT-2M, Synthbuster (fake images; average across generators) | accuracy: 60.5 -> 63.1 | Helped |
| Select the highest-scoring patch using the accuracy regression model instead of the center patch | End-to-end AIGI detectors · GenImage, DRCT-2M, Synthbuster (fake images; average across generators) | accuracy: 62.8 -> 63.1 | No clear effect |
| Select the lowest-scoring patch using the accuracy regression model instead of the highest-scoring patch | End-to-end AIGI detectors · GenImage, DRCT-2M, Synthbuster (fake images; average across generators) | accuracy: 63.1 -> 56.6 | Hurt |
| Select the highest-scoring patch using the ImageReward regression model instead of a random patch | End-to-end AIGI detectors · GenImage, DRCT-2M, Synthbuster (fake images; average across generators) | accuracy: 60.5 -> 61.8 | Helped |
| Select the highest-scoring patch using the HPS v2 regression model instead of a random patch | End-to-end AIGI detectors · GenImage, DRCT-2M, Synthbuster (fake images; average across generators) | accuracy: 60.5 -> 62.5 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | RINE · All generators in Table 3 | average precision: 90.2 -> 91.9 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | CoDE · All generators in Table 3 | average precision: 78.8 -> 79.2 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | SuSy · All generators in Table 3 | average precision: 87.6 -> 89.4 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | DRCT (Conv-B) · All generators in Table 3 | average precision: 84.8 -> 87.6 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | DRCT (CLIP) · All generators in Table 3 | average precision: 88.4 -> 86.7 | Hurt |
| Select the lowest predicted-accuracy patch instead of the center patch | NPR · All generators in Table 3 | average precision: 75.9 -> 78.6 | Helped |
| Select the highest predicted-accuracy patch instead of the center patch | NPR · All generators in Table 3 | average precision: 75.9 -> 74.6 | Hurt |
Surprising findings
- Higher-quality AIGIs are easier to detect for existing detectors, the reverse of the intuitive expectation.
- The easiest patches for end-to-end detectors tend to be the hardest patches for fingerprint-based detectors, and vice versa.
Limitations
- Quality is measured only by human preference models (ImageReward, HPS v2), not human judgment directly.
- Regression shows correlations; intervention experiments reveal non-uniform causal effects across detectors, so confounders remain.
- Results are based on a limited set of four generators and six detectors.
Experimental setup
Datasets: GenImage DRCT-2M Synthbuster CC3M LAION COCO
Benchmarks: GenImage DRCT-2M Synthbuster
Metrics: accuracy average precision accuracy on fake images accuracy on real images R2
Baselines: NPR RINE CoDE DRCT SuSy SSP
Builds on: ImageReward HPS v2 DRCT NPR SSP
Closest papers by meaning
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.89)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.88)
- A-Bench: Are LMMs Masters at Evaluating AI-generated Images? (ICLR 2025 · similarity 0.88)
- AI-generated Image Quality Assessment in Visual Communication (AAAI 2025 · similarity 0.88)
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.86)
- Semi-Truths: A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image detectors (NeurIPS 2024 · 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.