Research map / Image generation / Autoregressive and latent generation
Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now
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 generated images carry geometric defects by training classifiers on derived geometry cues (object-shadow pairs, perspective fields, line segments) that never see pixels, after prequalifying generated images to fool a pixel-based ResNet-50.
Problem
Generated images look realistic to people and to pixel-level detectors, but may still violate projective geometry (vanishing points, perspective, shadows); the paper asks whether this can be detected reliably from geometric features alone.
Approach
Prequalify generated images so that a pixel-based ResNet-50 is fooled, then train classifiers that see only derived geometry (object-shadow masks, perspective fields, line segments via PointNet) to separate real from generated images, so detection relies on geometric errors rather than signal artifacts.
Key results
- Geometry classifiers reach AUCs from 0.72 to 0.97 on prequalified test sets, with the prequalifier itself at chance (AUC 0.51) on the unconfident set.
- On the unconfident indoor set, the line-segment classifier reaches AUC 0.82, and the perspective-field classifier reaches 0.84 on the outdoor set.
- Kandinsky-trained classifiers generalize to SDXL, PixArt-α, DALL-E 3 and Adobe Firefly images, but not reliably to DeepFloyd.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use line segment cues rather than pixel signals to classify images | Kandinsky-v3-trained classifier · Indoor unconfident test set | AUC: 0.51 -> 0.82 | Helped |
| Use perspective field cues rather than pixel signals to classify images | Kandinsky-v3-trained classifier · Outdoor unconfident test set | AUC: 0.51 -> 0.84 | Helped |
| Use perspective field and object-shadow cues rather than pixel signals to classify images | Kandinsky-v3-trained classifier · Combined unconfident test set | AUC: 0.49 -> 0.80 | Helped |
| Train on Kandinsky-v3 and test on other generators | Geometry cue classifiers · Indoor and outdoor images from unseen generators | AUC | Mixed |
Surprising findings
- The Kandinsky-trained prequalifier generalized better than prequalifiers trained on other generators, and DeepFloyd images were hard to detect from models trained on other generators.
- Classifiers that never see pixels detect images that a pixel-based ResNet-50 cannot separate (AUC 0.51 on the unconfident set).
Limitations
- Classifiers were trained on prequalified data only where prequalification is done with Kandinsky images, and performance estimates are biased low as stated by authors.
- Generalization is uneven: DeepFloyd images are not reliably detected by Kandinsky-trained models.
- Analysis is limited to vanishing points and object-shadow cues; other geometry errors (scale, distortion, depth cues) are not evaluated.
Experimental setup
Datasets: LSUN Berkeley Deep Drive 100K Mapillary Vistas Kandinsky-v3 generated images Stable Diffusion XL generated images DeepFloyd IF generated images PixArt-α generated images recent social media images (May 2023 to March 2024)
Benchmarks: Kandinsky-v3 easy/unconfident/misclassified test sets (indoor, outdoor, combined) recent timestamp real/generated test set out-of-distribution generators (SDXL, DeepFloyd, PixArt-α, DALL-E 3, Adobe Firefly)
Metrics: AUC accuracy ROC curves Grad-CAM saliency
Baselines: ResNet-50 pixel prequalifier SOTA local signal-based detectors
Builds on: Perspective Fields Deep LSD PointNet SSISv2 ResNet
Closest papers by meaning
- Detecting Generated Images by Fitting Natural Image Distributions (NeurIPS 2025 · similarity 0.86)
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.84)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.84)
- Anomaly Score: Evaluating Generative Models and Individual Generated Images based on Complexity and Vulnerability (CVPR 2024 · similarity 0.84)
- Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.84)
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.