reference-analysis-validator

Reference Analysis Validator

This skill converts “looks close” into measurable gates. For brand/logo/mascot work, do not model or export until a source manifest and validation thresholds exist.

Required outputs

Create these in the asset output folder:

• reference_manifest.json — classified source files, expected parts, thresholds.
• source_analysis/*.json — image metadata, masks/components/landmarks.
• validation/front_overlay_reference.png — reference and render overlay.
• validation/front_mask_validation.json — IoU/SSIM/bbox/centroid report.

Workflow

1. Classify sources: front, side, back, top, texture atlas, decals, maps, lightmap, aura/context.
2. Build/refresh reference_manifest.json with hard expected counts and view roles.
3. Extract masks/components from each source using scripts/reference_manifest_compiler.py or existing analyzers.
4. Render model from matching orthographic camera with reference planes hidden.
5. Compare reference mask vs render mask using scripts/render_overlay_validator.py.
6. Refuse final export if hard gates fail.

Modality rule

Compare like with like. A wireframe edge mask compared against a shaded beauty render gives misleadingly low IoU. For hard gates, render a flat silhouette/matte pass from Blender or compare reference edges to render edges. Use render_overlay_validator.py --reference-mode ... --render-mode ... when the source and render need different mask extraction modes.

Default validation gates

• primary structural part count: exact.
• front silhouette IoU: target >= 0.90 for rigid/logotype shapes; >= 0.82 acceptable for first mascot reconstruction pass.
• bbox center drift: <= 12 px at 1024 px validation size.
• bbox size drift: <= 3% of image dimension.
• face/eye/smile landmark drift: <= 2% of image dimension when landmarks are defined.

Failure policy

If a repeated mismatch occurs, record the measured failure, then route to the missing specialty skill:

• wrong silhouette → contour-to-mesh
• wrong depth/side/back → orthographic-registration
• wrong textures → atlas-uv-fitting
• wrong whole workflow → mascot-logo-reconstruction

Read when needed

• references/metrics-and-thresholds.md for metric definitions and recommended gates.

Sources distilled

Official/library docs to prefer while extending this skill:

• OpenCV contour features: moments, area, perimeter, bounding rectangles.
• OpenCV shape matching / Hu moments.
• OpenCV homography and geometric transforms.
• scikit-image SSIM for perceptual comparison.