Skip to content

README.md

Latest commit

 

History

History
46 lines (38 loc) · 2.41 KB

README.md

File metadata and controls

46 lines (38 loc) · 2.41 KB
Raw

Reproducible rig review

rig_workflow.py generates a small textured triangle and two independent deformation bindings. It needs no private model, image files or Python packages beyond the Catchlight SDK. Run from the repository root:

nix develop -c cargo build -p catchlight-editor-server -p catchlight-cli
nix develop -c env PYTHONPATH=python python python/examples/rig_workflow.py \
  --cli target/debug/catchlight-cli --out /tmp/catchlight-rig-review

Use a fresh output directory. --server PATH optionally selects the editor server binary. The Nix shell supplies lavapipe when no hardware GPU is present.

The example composes ordinary operations:

  1. Create a synthetic model with binding-owned grids, an unset middle cell, texture, named vertex slots and a weld. Capture its revision and fork it.
  2. Read mesh and authored cells at that revision. Choose a cut through two triangle edges, map the new Part's offsets in client code, and supply exact deform_mapping weights for the retained Part. Refill its slots and add seam welds in the same atomic edit. Authored holes stay unset.
  3. Move a shared affine X deformation into a MeshGroup. The remaining local shape changes only Y, so these particular fields compose exactly. The example does not infer an exact transfer for arbitrary rigs; topology, representability and any fitting policy belong to the caller.
  4. Compare geometry at five chosen poses, export baseline/cut/assembled CLMs, and render each with the independent file CLI. Inspect each run.json, PNG and geometry sidecar. edits.json contains the explicit replayable batch.
  5. Replay that batch on the source at the captured revision. Its exported bytes must equal the reviewed candidate. Demonstrate that a second replay with the old guard is refused, then close the synthetic sessions.

The client's acceptance policy is an absolute geometry tolerance of 1e-5 world units at those poses plus successful rendering. Pixel comparison and broader pose selection can be added by the caller. report.json records the result without filesystem paths or machine identity.

python/tests/test_rig_workflow.py tests the same recipes without a GPU. It checks sparse cells, binding grids, slots, weld order, geometry, whole-batch rollback, exact candidate replay, stale guards and one-step Undo restoration.

nix develop -c uv run --project python pytest python/tests/test_rig_workflow.py -q