Skip to content

Captured mesh authoring

Implementation contract for F5, tracked by #4380. The viewer exposes captured surface creation inside Author → Appearance → Create from scan.

User workflow

Load a textured capture through the existing Open/Add path. From its Appearance workspace choose Create from scan, align the capture in the workspace, select a bounded region, and preview the resulting surface. Choose the destination IFC model and spatial container, then enter its Name. The initial semantic class is IfcBuildingElementProxy: captured surface geometry does not establish wall, slab, opening, material, or structural semantics.

Create publishes IFC entities, geometry, image ownership, containment and history as one transaction. The source remains independently available. Cancel or stale registration leaves no authored geometry or entities. Undo clears selection of the removed object; Redo restores its identity and appearance.

Canonical boundaries

  • Capture files enter useIfcLoader.loadFile(file, target). A capture-specific panel must not introduce another model-loading path.
  • Decode source geometry and retain the encoded image resources with their source lineage. Registration remains separate from file coordinates and from the chosen destination model's placement.
  • The native captured-mesh planner receives bounded IFC world metre coordinates, geometry indices, UV coordinates and independent UV indices. The adapter converts source coordinates and UV origin conventions exactly once. UV seams must survive even where several UV vertices share one geometry vertex.
  • Native planning resolves destination units, inverse spatial placement, IFC schema attributes and entity allocation. Reuse the existing appearance entity plan and produce_element_meshes path rather than constructing STEP text or a second preview tessellation in JavaScript.
  • The viewer transaction reuses the detached textured-owner insertion, source guards, asset inventory, containment helpers and history used by annotation creation. Multi-part geometry must retain one IFC object identity.

Input fidelity prerequisite

The GLB path converges through parseGlbViewerModel. The capture decoder preserves source UVs, original embedded albedo PNG/JPEG images, and node matrix or translation/rotation/scale transforms. Unsupported PBR maps and alpha modes are explicitly refused; see capture evidence.

Extend that canonical decoder or replace it in the same path; do not hide this gap by requiring users to convert their captures to IFC first. Unsupported material, UV or transform features must produce an explicit diagnostic before creation. A textured viewport image alone cannot certify that source UVs, encoded images and registration survive the authoring boundary.

The existing offline OBJ/MTL boulder adapter and IfcOpenShell writer are oracle experiments. They provide licensed source fixtures and independent comparison; they are not the production IFC implementation.

Acceptance

Use the real licensed public boulder capture and retain its source digest and license. Inspect a cropped result at a UV seam and compare source image samples. Verify actual hierarchy and viewport selection in one- and multi-model scenes, Name/class and spatial containment, cancellation and stale-result rejection, Undo/Redo, ordinary and subset IFCZIP, and fresh shared-room viewing. Validate exported IFC using an independent reader. Record resource limits and observed preparation/cancellation latency; do not claim reconstruction or watertightness from creation of an existing captured surface.

F4 transfer accuracy and F9/F10 reconstruction retain their independent real-data and quality gates. They are not prerequisites for creating an existing textured mesh region.

Shared host implementation

prepareAuthoredProduct serializes the effective IFC mutation view and captures its allocation/revision guard once for both annotations and captured surfaces. commitTexturedProduct owns publication, containment, original-image retention, selection cleanup on Undo and stable identity on Redo. Drawing creation uses the same transaction; there is no independent scan STEP writer.

prepareCapturedRegion binds a selected triangle region to its registered raw model geometry and one unambiguous retained original image. It preserves source sampling flags and converts GPU V coordinates once. createIfcFromCapturedMesh owns a bounded numeric snapshot before its first await, converts workspace coordinates into the destination model frame, and sends that snapshot to the native planner. It revalidates registration and target state before publication.

The integrated capture panel uses an independent instance of the existing renderer to preview the retained original image and selected source triangles. Orbit and zoom affect only that preview. A through-surface screen rectangle selects whole triangles by projected centroid; it never synthesizes border cuts. Preview graphics failure disables Create and offers a local preview reload. The original model, image/PDF draft controls, and main camera remain independent.

Actual WebGPU creation, original-image export, normal re-import, and picking are recorded in captured UI acceptance.

Destination eligibility is independent of load order or prior object selection: a loaded IFC4/IFC4X3 spatial model is offered even before its lazy mutation view exists. Creation initializes or reuses the canonical view, preserving prior edits. The created object activates its destination model so the shared Undo action targets that creation. Drawing-to-IfcAnnotation creation uses the same destination and selection helpers.

The native shared creator obtains metadata validation, the effective source context, allocation preflight and container-relative placement from appearance/authored.rs. Texture-specific row construction and canonical mesh production remain in the existing annotation/captured planner. This extraction does not change the public creator APIs or enable additional representations.