Blender verification tools
I built Blender tools that help an AI agent inspect the model it is making. The agent can render a view, compare it with a reference and examine the geometry behind a visible mismatch, then use those observations to guide another pass.
Left: reference image. Right: candidate modeled by GPT-6 Astra.
Silhouettes, selected regions and landmarks make proportion problems easier to locate. Scene bounds, mesh topology and camera projections add another kind of evidence: a shape can look plausible in one render while its dimensions or construction are wrong.
The comparison helped identify camera and proportion differences. View A's raw silhouette overlap rose from 0.69 to 0.90 across revisions. The validation record explains the segmentation assumptions, second-view result and remaining differences.
The tools connect modeling, rendering and inspection in one workflow. A comparison is always tied to its view and reference; I do not treat a numerical match as a universal quality score. The purpose is better feedback during iteration, especially when 'looks close' is hiding a specific error.