---
name: image-analysis
description: Combine visual, pixel, calibrated multi-view, and Blender evidence. Use for concept-to-model review, image-plus-mesh comparison, clipping diagnosis, or questions that must separate appearance from geometry.
---

# Combined image analysis

Use this skill when one answer must synthesize more than one evidence layer.
Keep each layer's claim boundary visible instead of turning unrelated signals
into a single confidence score.

## Route only what the task needs

- Pixel/file facts, exact colours, regions, or a two-image comparison: use
  [`image-measurement`](../image-measurement/SKILL.md).
- An existing template fit with calibrated projections, correspondences, and
  scale anchors: use
  [`multiview-reconstruction`](../multiview-reconstruction/SKILL.md).
- Blender mesh defects, depth-revealing renders, which part is in front, or a
  dent the flat renders hide: use the `blender-inspect` plugin's
  `model-inspection` skill (and its `blender-model-inspector` agent), when that
  plugin is installed. If it is not, say so; do not improvise mesh checks.
- Relative depth of a concept image, or whether it agrees with a model's depth
  render: `pil_depth.py` (`estimate`, then `compare`). See the depth row below.
- Image-plus-mesh review, Blender bounds, clipping or clearance, or a
  concept-to-model loop: continue here and load only the relevant sibling skill.

Several images do not by themselves justify a 3D solve. Without the required
camera and correspondence data, complete the valid 2D analysis and report the
reconstruction layer as underconstrained.

## Evidence layers

| Layer | Establishes | Does not establish |
|---|---|---|
| Visual inspection | Subject, style, apparent silhouette, likely problem areas | Exact pixel values, hidden geometry, collision |
| Image measurement | File facts, colour, structure, masks, local changes, projected silhouette | Metric depth, topology, rig or cloth behavior |
| Calibrated reconstruction | Constraints on an existing template and per-view residuals | Unseen topology or meaning from pixels |
| Blender scene probes (`blender-inspect`) | Mesh counts, bounds, topology defects with locations, depth order, signed clearance | Visual or semantic match by itself |
| Concept-image depth (`pil_depth`) | Model-inferred *relative* inverse depth of one image, and its rank agreement with a render's depth | Metric depth, hidden geometry, or which source is right where they disagree |
| Locked render review | Matched projected appearance across requested views | Physical clearance or deformation quality |

## Working method

Inventory and match the requested views before comparing them. Inspect them
visually before reading scores, then use `image-measurement` for reproducible
pixel/file evidence. Keep normalization, masks, registration transforms, raw
measurements, and domain-profile identity with the result.

Prepare every requested multi-view contour, including refused views. Run a
template solve only when calibration and correspondences are sufficient, and
continue to geometry mutation only from `SOLVED`. Use
`blender_fit.py --mode probe` for penetration or clearance,
`blender_mesh.py` for counts and bounds, and `blender_mesh_audit.py` for
topology defects, all from `blender-inspect` (load its `model-inspection`
skill, which resolves that plugin's scripts directory); silhouettes are not
collision evidence. The `pil_blender_*` copies of these tools are deprecated and go away
in the next minor release: use them only when `blender-inspect` is not
installed, and say that the deprecated copy ran.

Concept depth is one extra layer, not a verdict: `pil_depth.py estimate`
gives model-inferred relative inverse depth, and `compare` aligns it to a
render's depth `.npy` from `blender_inspect_render.py`. Read the Spearman rank
correlation together with AbsRel, `valid_pixels`, the region boxes and the
disagreement image; AbsRel alone tracks camera distance, and the two views need
matching framing and pose. There are no pass/fail thresholds.

Render only the decisive requested views with locked framing, then aggregate
matched comparisons with `pil_multiview_review.py`. A missing, refused, or
violated required view must remain visible in the overall result.

Run `pil_capabilities.py` to list every tool's arguments and dependency
status, or `pil_capabilities.py --tool pil_palette_diff` for a single tool.
The executable 0.9.0 normalization, selection, registration, local-diff,
calibration, batch, and MCP workflows are in [measurement workflows](../../docs/measurement-workflows.md).

## Report and finish

Group findings by evidence layer. For a requested match target, define the
required views, metrics, thresholds, geometry constraints, and worst-case rule
before judging it. Report the weakest required view, all refusal states, and
the assumptions that limit each claim. The analysis is complete when every
requested layer is measured or explicitly marked unavailable; do not stop at a
first render or first plausible score when the request includes the full loop.