PanorAi 3.2.0

PanorAi 3.2.0 adds an Experimental ergonomic workflow over the stable panorai.geometry foundation while preserving the 3.x API.

Highlights

• Start from panorai.EquirectangularImage and attach radial depth or integer labels with explicit validity.
• Generate deterministic cube, Fibonacci, icosahedron, or spiral views without manually assembling channels or projection plans.
• Apply array or tensor models through views.map() and reconstruct with modality-safe interpolation and blending.
• Use process_views() as the exact shorthand for views().map().reconstruct().
• Inspect the executed geometry, modalities, interpolation, validity, backend, device, and reconstruction policy with views.describe().
• NumPy HW/HWC and Torch HW/CHW/NCHW workflows preserve backend, batch, dtype, device, support, validity, and autograd.

Semantics and compatibility

The new surface is Experimental. Existing 3.x names and advanced APIs remain available. Image and radial depth are continuous and bilinear; labels are nearest-only. Strict invalid-data propagation remains the default, with explicit mask-authoritative renormalization available as an opt-in policy.

The distribution retains lightweight adapters for optional depth backends but does not vendor model implementations, datasets, checkpoints, or training trees.

The protected release workflow builds one wheel and one sdist, audits and attests them, verifies byte identity through TestPyPI, runs installed NumPy/Torch geometry and documentation checks, and promotes the same files to PyPI only after protected approval.

Full details: https://github.com/rlsgarcia-code/PanorAi/blob/v3.2.0/CHANGELOG.md