Comparison Algorithm

imugi uses a multi-signal comparison approach to objectively measure how close your code is to the design.

Signals

Signal	What it measures	Weight
SSIM	Structural similarity — luminance, contrast, structure	0.4 (with vision) or 1.0 (alone)
Pixel Diff	Raw pixel-level comparison via pixelmatch	Used for heatmap
Claude Vision	AI-powered visual assessment	0.6 (when available)

Composite Score

The composite score is calculated as:

• SSIM only (when vision unavailable): score = ssim
• SSIM + Vision: score = ssim * 0.4 + vision * 0.6

The score is clamped to [0, 1].

When Vision is Used

Claude Vision analysis is activated when SSIM < 0.98. For very high SSIM scores, pixel-level comparison is sufficient and the API call is skipped.

Diff Regions

Pixel differences are clustered into regions using connected component analysis. Each region gets:

• Position (x, y) and size (width, height)
• Diff intensity — average difference magnitude
• Pixel count — number of differing pixels

Region Classification

Regions are classified by type:

Classification	Description
color	Color mismatch
spacing	Margin/padding difference
size	Element size mismatch
position	Element position offset
missing	Element present in design but not in code
extra	Element in code but not in design
font	Typography difference

Priority Assignment

Priority	Criteria
high	Large area, high intensity, structural differences
medium	Moderate area or intensity
low	Small area, low intensity

Strategy Selection

Based on the composite score, imugi selects a patching strategy:

Score	Strategy	Description
< 0.7	Full regeneration	Rewrite the code from scratch
>= 0.7	Surgical patch	Targeted fixes for specific regions

The strategy can also switch based on iteration history — if surgical patches are not improving the score, imugi may escalate to full regeneration.