Statistical Analysis of Scene Structure in Real vs Synthetic Images

This project investigates how real photographs differ from AI-generated images (Stable Diffusion v1.5) by analyzing their geometric structure and frequency content using classical computer vision methods. The goal is not to build a classifier, but to understand where and how generative models diverge from real-world scene statistics.

Dataset

We study four scene categories:

• Forests
• Highways
• Cockpits
• Bedrooms

Real images come from:

• Toronto Scenes dataset [11]
• SUN397 dataset [13]

Synthetic images (~1,000 per category) were generated with Stable Diffusion v1.5 using prompt engineering (base scene description, lighting, key objects) plus negative prompts to avoid artifacts. All images were resized/filtered to a common resolution of 800×600, converted to grayscale, normalized, and smoothed with a 5×5 Gaussian kernel.

Methods

We use two complementary perspectives:

1. Geometric Analysis (Contour-Based)

Using OpenCV, we:

• Extract edges (Canny) and external contours (findContours).
• Compute contour length statistics (mean, median, std) as proxy for structural complexity.
• Estimate contour orientations via ellipse fitting (fitEllipse) and build orientation histograms.
• Measure parallelism: fraction of contour pairs whose orientations differ by < 10° and lie within a distance threshold.
• Estimate junction density: locations where multiple contours meet within a small radius, normalized by image size.

These features characterize scene structure in terms of shape complexity, directionality, and how often structures intersect or align.

2. Frequency Analysis (DCT-Based)

To analyze texture and structural regularity, we apply the 2D Discrete Cosine Transform (DCT):

$$ D(u,v) = \frac{1}{\sqrt{2N}} \sum_{x=0}^{N-1}\sum_{y=0}^{N-1} I(x,y)\cos\left(\frac{\pi(2x+1)u}{2N}\right) \cos\left(\frac{\pi(2y+1)v}{2N}\right) $$

We:

• Compute average DCT spectra for real vs synthetic images.
• Visualize difference heatmaps in the frequency plane.
• Compare histograms of DCT coefficient magnitudes.
• Aggregate energy in low / mid / high frequency bands to see how power is distributed.

This reveals whether generative models introduce periodic artifacts or distort the natural frequency distribution of scenes.

Key Findings

Geometric (Contour) Structure

• Forests show the largest gap between real and synthetic:
  • Synthetic forests tend to have shorter contours and less variance, hinting at oversimplified vegetation.
  • Junction densities are often higher in synthetic forests, suggesting overcompensated fine detail (e.g., too many small intersecting structures).
• Highways show moderate differences:
  • Real highways exhibit broader contour length distributions and higher junction densities.
  • Synthetic highways often have lower parallelism scores and fewer complex intersections, implying simplified infrastructure.
• Cockpits are closest between real and synthetic:
  • The strong geometric constraints of man-made interiors appear easier for the model to replicate.
• Bedrooms show an interesting flip:
  • Synthetic bedrooms often have longer contours and more clutter, possibly due to the model adding extra objects.
  • Junction densities in synthetic bedrooms are sharply peaked, suggesting “template-like” structures (e.g., many very similar bed layouts).

Overall, synthetic images frequently either oversimplify natural complexity or overcomplicate certain details, deviating from real structural statistics.

Frequency Domain (DCT) Structure

• Difference heatmaps show that synthetic images have stronger low-frequency coefficients, emphasizing smoother, more homogeneous patterns.
• Real images tend to have more energy in higher frequency regions, reflecting natural fine details and sharper edges.
• Histograms of DCT coefficients:
  • Real images have a higher probability of near-zero coefficients, consistent with large uniform regions (e.g., sky, walls).
  • Synthetic images show broader coefficient distributions, suggesting added noise and less “sparse” frequency structure.
• Band-wise analysis (low/mid/high) indicates that synthetics often have elevated energy across all bands, while real images show a more natural, decaying spectrum.

These results suggest that current generative models do not replicate the natural distribution of spatial frequencies and instead produce images with more diffuse energy across the spectrum.

Conclusion & Future Directions

This work shows that classical computer vision tools—contours, orientations, junctions, and DCT—can expose systematic differences between real and AI-generated images that aren’t obvious to the naked eye. Even using an older model (Stable Diffusion v1.5), we find:

• Geometric statistics (contours, parallelism, junctions) differ meaningfully, especially in complex natural scenes.
• Frequency statistics (DCT patterns) reveal non-natural energy distributions in synthetic images.

Limitations and next steps:

• This study used a single model (SD v1.5) and four scene categories; extending to modern models and more categories would make the conclusions stronger.
• More rigorous prompt protocols and statistical tests (e.g., distribution distances, hypothesis testing) could quantify differences more precisely.
• Future work could add color statistics, texture descriptors, curvature, symmetry, and other higher-order features, and track how these statistics evolve with newer generative models.