Research map / Image generation / Autoregressive and latent generation
Stay-Positive: A Case for Ignoring Real Image Features in Fake 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 Stay-Positive, which retrains only the last layer of a fake-image detector with non-negative weights so that real-associated features cannot push predictions toward real, improving generalization to FLUX/aMUSEd and robustness to WEBP/downsizing.
Problem
Fake image detectors trained on known generator families learn spurious correlations with real-image features (e.g., WEBP compression in real training data, image quality), which hurts generalization to newer generators such as FLUX and robustness to post-processing, and makes partially inpainted re
Approach
Under a ReLU features-plus-linear-head assumption, negative last-layer weights correspond to real-image features; retraining the last layer with weights clamped to be non-negative (with the backbone frozen) forces the detector to decide using only fake-associated features.
Key results
- On the custom benchmark, Corvi⊕ raises FLUX AP from 57.25 to 99.33 and Rajan⊕ raises FLUX AP from 80.64 to 90.50 (no post-processing), with average AP 99.88 vs 94.23 for Corvi.
- On partially inpainted images at 50% inpainted pixels, Corvi AP is 6.13 while Corvi⊕ reaches 97.48; Rajan goes from 49.69 to 99.66.
- On GenImage VQDM, Corvi⊕ AP improves from 98.68 to 99.94 and Rajan⊕ from 95.11 to 99.26.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Retrain the last layer with non-negative weights and a frozen backbone | Corvi · All generators (AVG) | AP: 92.80 -> 99.88 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Rajan · All generators (AVG) | AP: 96.22 -> 98.65 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Corvi · FLUX | AP: 57.25 -> 99.33 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Rajan · FLUX | AP: 80.64 -> 90.50 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Corvi · 50% inpainted images | AP: 6.13 -> 97.48 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Rajan · 50% inpainted images | AP: 49.69 -> 99.66 | Helped |
| Clamp final-layer weights to non-negative values without retraining (ablation) | Corvi · All generators (AVG) | AP: 92.80 -> 80.47 | Hurt |
| Clamp final-layer weights to non-negative values without retraining (ablation) | Rajan · All generators (AVG) | AP: 96.73 -> 88.17 | Hurt |
| Retrain the whole network while clamping final-layer weights, instead of freezing the backbone (ablation) | Corvi · FLUX | AP: 94.23 -> 69.50 | Hurt |
| Retrain the whole network while clamping final-layer weights, instead of freezing the backbone (ablation) | Rajan · FLUX | AP: 91.85 -> 58.68 | Hurt |
| Retrain the last layer with non-negative weights and a frozen backbone | Corvi · VQDM (GenImage) | AP: 98.68 -> 99.94 | Helped |
| Retrain the last layer with non-negative weights and a frozen backbone | Rajan · VQDM (GenImage) | AP: 95.11 -> 99.26 | Helped |
Surprising findings
- Clamping weights without retraining the backbone degrades performance substantially (Corvi AP average 80.47 vs 92.80), and retraining the whole network with clamping underperforms on FLUX.
- The method can still inherit spurious fake features: Corvi⊕ continues to classify upsampled real images as fake.
Limitations
- Fake-distribution spurious features (e.g., upsampling) are not addressed and persist in the method.
- Real-specific features can still implicitly shape fake decisions via negation, so full-network training may be needed.
- Relies on assumptions (ReLU before final linear layer, sigmoid output) for identifying real-associated features.
Experimental setup
Datasets: COCO LSUN Redcaps WikiArt LAION-Aesthetics whichfaceisreal InstructPix2Pix Nights MidJourney v4/v5 Kandinsky 2.1 Playground v2.5 PixArt-alpha
Benchmarks: GenImage UFD benchmark Synthbuster CNN-generated images test set (Wang et al.) custom 9-generator post-processed test set
Metrics: average precision accuracy logit score
Baselines: Corvi Rajan AEROBLADE UFD-ProGAN UFD-LDM ClipDet DRCT GAN-Baseline Cozzolino-LDM
Builds on: Corvi detector Rajan detector (LDM reconstruction training) UFD ResNet-50
Closest papers by meaning
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.89)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.88)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.87)
- Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.87)
- Spot the Fake: Large Multimodal Model-Based Synthetic Image Detection with Artifact Explanation (NeurIPS 2025 · similarity 0.87)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · 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.