Clean up AI images. Make text and lines clearer.

After loading, select flat patches at original resolution and calculate the standard deviation of luminance residuals. Contour width and text clarity are analyzed separately.

σ is measured in 8-bit intensity levels (0–255), not a 0–100 quality score. Lower values mean less fine luminance variation in the selected areas, not necessarily a better overall image.

This method may mistake intentional texture for noise. It does not cover purely chromatic noise or large blotches. No score is shown when too few patches qualify.

Measure 10–90% widths of stable image contours at original resolution, in px.

Measures the pixel distance for a contour to move from 10% to 90% of its local intensity difference. Smaller means a steeper transition, not a thinner object, a 0–100 quality score, or test-chart MTF. This scans general contours, including possible text; text clarity uses text candidates only.

Uses high-contrast edges with stable plateaus on both sides and excludes profiles touching transparency. At least 32 valid profiles are required. Before and after use fixed locations, directions, and plateau ranges, with 10%/90% levels derived from each image's plateaus. If any profile loses correspondence, comparison is unavailable; samples are not reselected.

Edge narrowing is independent and on by default at 70%. If earlier processing makes profiles unmatchable, narrowing is skipped with a reason. To inspect its effect alone, turn off text sharpening and image detail enhancement.

Resolution, angle, sampling, and sharpening halos affect the value. Narrower edges do not guarantee freedom from jaggies, broken lines, or halos. Inspect the image as well.

Measure Sobel gradient energy in fully opaque neighborhoods at original resolution.

Higher Tenengrad energy means stronger intensity changes, not a 0–100 quality score. Compare only the same image at the same size. Noise, contrast, and texture can also raise it; a decrease after denoising does not necessarily mean blur. Check noise, text metrics, and the image together.

Image detail enhancement is optional and runs after denoising, before text sharpening. Only detected text candidates are protected; missed text may be processed. This feature takes longer and can be turned off.

Excludes the outer border and 3×3 neighborhoods touching transparent or translucent pixels. No score is shown with fewer than 16 valid positions or less than 1% coverage. Zero is valid for a flat-color image and means no intensity gradient.

Isolate stroke cores in dark text candidates and calculate CIEDE2000 color difference RMS and P90 against each group's original dominant color.

Lower values mean selected core pixels are closer to their group's reference color. Core positions and target colors stay fixed before and after. Antialiased edges are excluded. This is not text accuracy or a universal cleanliness score. No value is shown with insufficient samples.

Supports dark, nearly monochrome horizontal text on light backgrounds. Suspected colored or widely varying text is skipped, but graphics may still be mistaken for text and thin strokes may be missed. At 100%, eligible cores converge to each group's original dominant color; changes are reverted if stroke edges are affected. Edges remain unchanged, so some color variation may remain.