Case study
Applied AI
Structural Verification for Generated Images
A deterministic quality gate for generative image edits, using a hand-implemented Sobel edge detector and edge-overlap IoU scoring to flag structural changes.
Outcome
Deterministic geometry checks around a non-deterministic staging model.
The constraint
Generative staging can produce a plausible room while quietly moving a wall, window, or door. Visual quality alone is not enough for a listing image; the pipeline needs a deterministic check that can flag structural drift for human review.
Verification path
01
Original
Immutable source image
02
Generated
Staged model output
03
Edge maps
Explicit 3×3 Sobel kernels
04
Overlap
Edge-intersection IoU
05
Gate
Pass, flag, or fail
Engineering decision
Put a deterministic check around the generative step.
- Convert both images to raw grayscale buffers and calculate horizontal and vertical gradients
- Threshold gradient magnitude into comparable structural edge maps
- Score the intersection over union of edge pixels instead of trusting a visual-quality model
- Keep the scoring code unit-tested and separate from the staging model
- Send borderline or conflicting results to review rather than silently accepting them
Staging output
Before
After
Before
AfterInternal production tool; access is restricted.
3×3
Sobel Kernels
IoU
Edge-Overlap Score
Fail Closed
Ambiguous Results
Tested
Deterministic Gate
Built With