Research map / Image generation / Autoregressive and latent generation
A Bias-Free Training Paradigm for More General AI-generated Image Detection
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 B-Free, a bias-free training paradigm that fine-tunes a DINOv2+registers ViT end-to-end on self-conditioned Stable Diffusion 2.1 reconstructions of COCO images with inpainting-based content augmentation, for generalizable and calibrated AI-generated image detection.
Problem
AI-generated image detectors perform well on supervised benchmarks but fail to generalize and calibrate on real-world images, largely because training data contains spurious correlations (content, format, resolution) rather than generator artifacts.
Approach
Generate fake images as self-conditioned diffusion reconstructions of real images (content-aligned via Stable Diffusion 2.1 inpainting with an empty mask), then add content-based inpainting augmentation and avoid resizing so detectors learn generator artifacts rather than dataset biases.
Key results
- B-Free achieves 98.0 AUC / 92.2 bAcc on the ablation average over Synthbuster, new generators and WildRF, improving from 93.5 bAcc for text-paired training.
- Across 27 generators, B-Free outperforms the second-best method by +20.7% balanced accuracy in the per-generator comparison (Table 5).
- On the viral-content analysis, B-Free keeps around 92% accuracy even many days after first online posting, while most competitors fall below 70%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use self-conditioned image generation rather than text-driven pairing for training data (ablation) | B-Free · Synthbuster and WildRF (AVG) | balanced accuracy: 61.9 -> 81.4 | Helped |
| add inpainting augmentation to self-conditioned generations (ablation) | B-Free · Synthbuster and WildRF (AVG) | balanced accuracy: 81.4 -> 92.2 | Helped |
| add same- and different-category inpainting plus standard augmentations (ablation) | B-Free · Synthbuster and WildRF (AVG) | balanced accuracy: 92.2 -> 95.2 -> 96.4 | Helped |
| apply cutmix/mixup instead of inpainting to self-conditioned generations (ablation) | B-Free · Synthbuster and WildRF (AVG) | balanced accuracy: 81.4 -> 78.6 | Hurt |
| fine-tune end-to-end rather than train only the final linear layer (ablation) | DINOv2+reg · Synthbuster and WildRF (AVG) | balanced accuracy: 68.5 -> 95.2 | Helped |
| add registers to DINOv2 (ablation) | DINOv2 · Synthbuster and WildRF (AVG) | balanced accuracy: 91.1 -> 95.2 | Helped |
| replace DINOv2+reg with SigLIP (ablation) | B-Free · Synthbuster and WildRF (AVG) | balanced accuracy: 95.2 -> 89.9 | Hurt |
| use the B-Free dataset to train CLIP/ViT instead of ProGAN or Latent Diffusion data (ablation) | CLIP/ViT · Synthbuster and WildRF (AVG) | balanced accuracy: 45.2 -> 73.9 | Helped |
| use the B-Free dataset to train RINE instead of ProGAN or Latent Diffusion data (ablation) | RINE · Synthbuster and WildRF (AVG) | balanced accuracy: 65.0 -> 83.2 | Helped |
| train B-Free with its proposed self-conditioned, inpainted++ data paradigm | AI-generated image detection · Synthbuster, new generators, and WildRF (AVG) | balanced accuracy: 75.6 -> 96.3 | Helped |
| train B-Free with its proposed self-conditioned, inpainted++ data paradigm | AI-generated image detection · GenImage unbiased, FakeInversion, and SynthWildX (AVG) | balanced accuracy: 74.2 -> 90.4 | Helped |
Surprising findings
- Linear probing of the frozen DINOv2+reg features performs poorly (68.5 bAcc) versus end-to-end fine-tuning (95.2), suggesting last-layer semantic features miss low-level artifacts.
- Several SoTA detectors (e.g., NPR, LGrad) show highly non-uniform performance across datasets, suggesting they exploit dataset biases.
Limitations
- Approach relies on fakes from Stable Diffusion 2.1; may fail for future generators with very different synthesis processes.
- As a data-driven approach, it can be adversarially attacked by malicious users.
- Evaluated on a limited set of generators; robustness to new architectures is not established.
Experimental setup
Datasets: MS-COCO (training split) Synthbuster GenImage (unbiased, modified) FakeInversion WildRF SynthWildX
Benchmarks: Synthbuster GenImage (unbiased) FakeInversion SynthWildX WildRF New Generators (Midjourney, SDXL, DALL-E 2/3, Firefly, FLUX, SD 3.5)
Metrics: AUC balanced accuracy balanced NLL ECE
Baselines: CNNDetect DMID LGrad UnivFD DeFake DIRE AntifakePrompt NPR FatFormer FasterThanLies RINE AIDE
Builds on: DINOv2 Stable Diffusion inpainting CutMix mixup CLIP linear probing (UnivFD)
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.92)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.90)
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.90)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.89)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.88)
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.87)
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.