Proof to Model

Proof to model

PROOF_TO_MODEL turns authorized original images into editable, measured 3D assets through source classification, six approved proofs, native multi-view reconstruction and Blender finishing. It is the default primary art pipeline, but selecting it does not generate views, provision models or approve runtime promotion.

Read the pipeline contract, coordinating skill and command argument reference before production. Schemas, recipes and fixed execution resources belong to the pinned Bubo dependency.

Prerequisites and state

Resolve STYLE and any task-local prompt selection, architecture, effective engine and compatible skills. Establish source rights, scale, intended format, geometry/texture/collision budgets and rig/animation applicability. If external libraries are configured, inspect them first and use only verified repository-local copies.

Initialize and build Bubo explicitly with an existing .NET 10 SDK before native operations. Setup is a separate dry-run-first operation requiring applicable license and pinned-file evidence before execution. Ordinary production does not download tools/models, install drivers or accept licenses.

Keep task state under .hephaera/.cache/tasks/<task>/, immutable originals and versioned attempts under .art-staging/proof-to-model/<asset>/<version>/, and audited dependencies under the declared ignored local-tool location. Reject linked escapes and runtime/staging overlap. These guards do not create an OS sandbox.

Register and classify original sources

init requires task, asset, version and actor identity, exactly one original --proof or a --sources JSON set, plus --rights, --recipe and --toolchain. Use current schemas and real project-confined files. Multi-view is the default; explicit single-view diagnostic mode is a separate nonproduction path.

The assistant inspects and researches the registered originals. classify --classification JSON registers the resulting report; it does not run vision. Record distinct entities, living anatomy, equipment, nonliving assemblies, attachment ownership, materials, pose, occlusion and asymmetry. Distinguish observed, research-inferred, user-specified and unknown claims. General research can support plausibility, not unseen source-specific detail.

Generated proofs and model renders never become original-source classification evidence, even after copying or renaming. Resolve consequential unknowns and conflicting originals before dependent production. Keep a rider distinct from a mount and equipment distinct from anatomy; intentional fantasy features should not be “corrected” without evidence.

Prepare exactly six consistent views

The ordered set is front, subject-right, back, subject-left, top and bottom. Freeze identity, pose, scale, assembly, materials and camera/frame convention. One usable canonical original can supply a slot and need five additional proofs; an oblique/cropped anchor may need all six canonical views. Multiple originals can reduce missing coverage.

prepare-proofs --proof-set JSON registers actual image/camera/provider artifacts. It does not generate images. Generated views need separately selected CLOUD or LOCAL production under MEDIA, explicit candidate/revision limits and provenance. Show originals and all six current views together, including inferred hidden surfaces, for human selection.

Use the current status proofSet fingerprint for the proof-set approval kind. A diagnostic single proof receipt cannot approve the set. Target revisions retain rejected candidates and unchanged approved bytes; any set revision invalidates whole-set approval.

Camera and preprocessing contract

The canonical rig uses row-major Blender/NeRF camera-to-world matrices: world Z up, camera local -Z forward and +Y image up. Front faces -Y, subject-right is -X and subject-left is +X. Preserve these labels across prompts, attachment review and ordered inference frames. Record scale, distance, horizontal FOV and explicit top/bottom roll. Per-view FOV overrides the shared value.

Prescribed views derived from an illustration are nominal-prescribed; do not invent measured calibration. Inspect actual alpha per approved image. Preserve suitable mattes or explicitly use pinned BiRefNet through Pixal3D. Retain masks, cutouts, common-square padding transforms and adjusted FOV without stretching or independently tight-cropping views. Actual preprocessing needs review; placeholder threshold matting is not a substitute.

Native reconstruction and finishing

Pinned Pixal3D MV inference must receive six separate verified prepared frames and camera JSON with the actual six-view request. A contact sheet, single-view donor or four-image subset cannot satisfy production. Retain submitted hashes, raw output, model/executable identities, logs and bounded GPU/process evidence. Upstream validation covers four views; accepting six inputs does not establish six-view quality. Separate bounded real acceptance is required.

Blender 4.5 LTS finishing uses only fixed audited hash-bound finish/review scripts on declared working copies. Preserve raw reconstruction and source high geometry. Apply a declared raw-engine frame correction once and record it.

Finishing route	Required evidence	Prohibited shortcut
Default QuadriFlow	Actual successful retopology before applicable baking	Claiming success from a requested operator or failed solver
Explicit voxel preparation	Separate low working copy and current finish settings	Altering the immutable high source
Explicit schemaVersion2 rigid-parts	Independently reviewed authored HIGH/LOW pairs, repair/part/pivot ancestry, budgets and native memory enforcement	Automatic route switch after solver failure or relabeling a proxy HIGH

Preserve valid UVs or unwrap with the declared margin. Applicable real high-to-low/grouped baking retains cage/ray/tangent settings and material correspondence. Measure textures, LODs and per-part collision. Explicit conservative kdop-26 collision needs containment, convexity, face and volume checks. Never loosen tolerances or expand budgets automatically.

Command stages and review receipts

Every continuation uses the exact current --manifest path returned by initialization:

bash .hephaera/commands/proof-to-model.sh status --manifest PATH_TO_MANIFEST
bash .hephaera/commands/proof-to-model.sh validate --manifest PATH_TO_MANIFEST

Operation	Purpose
doctor, setup	Inspect prerequisites; setup is dry-run without explicit execution
classify, prepare-proofs	Register actual reports and six-view artifacts
preprocess, reconstruct, finish	Run the current approved native stages
review	Produce measured comparison evidence
record	Consume an actual current receipt with kind, reviewer and evidence
reuse	Check donor ancestry after receiving-asset registration and review
validate, status	Inspect current contract/evidence state
promote	Copy exact accepted output to a declared fresh assets/models/... destination

finish --finish-settings JSON and --finish-budget JSON require their explicit current evidence. An exhausted finish budget may receive one separately approved additional attempt bound to consumed history; never reset the ledger.

Save editable Blender sources and GLB, independently re-import the export and inspect transforms, hierarchy, materials, textures, dependencies and budgets. Compare originals, proofs, prepared inputs and six matching-camera GLB renders with honest nominal-fit labels. Schema checks cannot establish likeness or anatomy.

Current independent review and separate human runtime acceptance must bind exact artifacts before promotion. Rigging/animation need subject-specific readiness and motion evidence. Historical v1/single-view diagnostics remain historical, cannot fall back into six-view production and cannot promote. Changed inputs or runtime builds require retained-ancestry versions and refreshed affected reviews; simulated process fixtures never become live reconstruction claims.