bsmi021/pil-agent-plugin

Combines native visual inspection, quantitative Pillow image measurement, calibrated multi-view reconstruction constraints, Blender BVH geometry checks, and matched-render review. Use when an image-analysis request spans several evidence layers, such as concept-to-model comparison, seven-view character or garment review, clipping diagnosis, or iterative image-and-mesh evaluation. Use image-measurement alone for pixel-only questions and multiview-reconstruction alone for an already-calibrated template fit.

Measures images numerically and compares two images with Pillow — exact hex colour palettes with coverage, per-hue census, luminance/saturation/entropy statistics, per-region grid statistics, perceptual hash distance, and changed-region bounding boxes. Use when asked whether two images match, whether a render or screenshot kept a reference's colour scheme or layout, what specifically changed between two versions, what the exact colours of an image are, or to compare a generated image against a reference. Does NOT measure 3D geometry or polygon count.

Prepares calibrated multi-view concept images, fits a template mesh with OpenCV and SciPy constraints, checks Blender garment clearance with BVH, renders arbitrary locked-framing views, and performs worst-case image review. Use for reconstructing or fitting an existing 3D mesh from several image views when camera projections, correspondences, and scale anchors can be supplied. Does not infer hidden topology or metric 3D geometry from pixels alone.