contour-to-mesh · git:20260921.d4bf893 · 2026-09-21 · sha256 3d5648094c36866d
contour-to-mesh git:20260921.d4bf893A
Immutable. This exact content is served forever at /api/v1/blob/3d5648094c36866d.
--- name: contour-to-mesh description: Build Blender mesh surfaces directly from extracted 2D contours/masks instead of approximate primitives. Use for 1:1 mascot/logo reconstruction, exact leaf/petal silhouettes, filled wireframe shapes, shallow bas-relief forms, or when the model outline must match a front template before adding depth. when_to_use: Source-locked contour-derived meshes, silhouette-first modeling, triangulated mask/contour surfaces, mesh generation from OpenCV contours, or replacing generic ellipses/primitives with measured shapes. allowed-tools: Read Bash Glob Grep mcp__blender__execute_blender_code mcp__blender__get_scene_info mcp__blender__get_object_info --- # Contour to Mesh Use this when a reference silhouette is the contract. The mesh starts as a filled 2D contour in front-view X/Z, then gets shallow Y depth/crown after validation. ## Workflow 1. Extract or select a clean contour/mask for one structural part. 2. Generate a triangulated mesh recipe with `scripts/mask_to_mesh_recipe.py`. 3. Run the generated Blender Python to create a named `GEO_*` mesh. 4. Assign front-projected UVs immediately. 5. Add depth only as a modifier or vertex displacement that preserves X/Z boundary positions. 6. Validate the front render against the source mask. ## Hard rules - Do not use radial duplication unless the manifest says the source is symmetric. - Do not use a generic ellipse if a contour exists. - Boundary vertices must keep source X/Z coordinates. - Depth deformation may move Y and inner vertices, but must not alter front silhouette. - One structural source part should become one named structural mesh. ## Blender pattern Generated mesh coordinates should map image `x` to Blender `X` and image `y` to Blender `Z` with vertical flip. Use `Y` only for thickness/depth. ## Front-plane rotation rule In this Blender coordinate convention, the front camera looks along the Y axis and the reference silhouette lives in the X/Z plane. Therefore 2D rotations inside the front view are rotations around the **Y axis**, not around Z. Use `rotation_euler = (0, angle, 0)` for 2D component orientation in the front projection. Z rotation spins objects into/out of screen-space incorrectly for X/Z-plane meshes. ## Sources distilled - Blender Mesh API: create meshes from vertices/faces using `from_pydata`. - Blender BMesh: use for cleanup, triangulation, normals, smoothing. - OpenCV contours: detect and simplify boundaries. - Delaunay triangulation: useful for filled interior meshes when filtered by mask containment. ## Part-inventory handoff When masks come from `source-part-segmentation`, preserve the `part_inventory.json` names in Blender: `GEO_<part_name>`, `MAT_<part_name>`, and UV layer names. Mesh generation must record: - source image size; - source mask path; - contour boundary point count; - boundary landmark points (tip, base-left, base-right, centroid/extrema) when detectable; - atlas UV rectangle or projection mode. For exact mascot/logo work, generate boundary vertices from the contour and add interior vertices only for surface support. The boundary is a contract: bevels, solidify, subdivision, shrinkwrap, or sculpt passes must not move front-view X/Z boundary vertices unless a new source mask is produced. ## Hole and stroke policy - Filled structural masks become meshes. - Wireframe strokes guide boundaries/landmarks; they are not structural meshes unless the manifest classifies them as decorative linework. - Holes/negative regions should be kept as separate cut masks when they affect silhouette; otherwise implement them as decals/material masks. ## Source-locked front-skin fallback When a design sheet is painterly/stylized and separate orthographic views are not CAD-consistent, use a **front-skin** pass before sculptural interpretation: 1. choose the canonical front/master source; 2. extract the visible subject mask as a filled contour; 3. create a shallow 2.5D mesh whose X/Z boundary is that contour; 4. assign full-image front-projected UVs so the rendered front matches the source pixels; 5. add depth only along Y and behind the front surface; 6. optionally add separate relief/backing parts for side plausibility, but keep the front skin as the visual acceptance gate. This is a generic fallback for logos/mascots where a faithful front read is more important than speculative volume. Mark it as `front_locked_visual_skin`, not as a fully solved turntable model. ## Script - `scripts/source_locked_skin_recipe.py` extracts a largest/canonical foreground component and emits a mask + contour mesh recipe using full-image front-projected UV coordinates.