Changelog

0.8.0 (unreleased)

Support for the new attribute types introduced in Blender 5.2, alongside a cleanup of the attribute typing system. Requires bpy >= 5.2.

Support for the new Blender 5.2 attribute types:
- FLOAT4 — 4D float vectors (Float4Attribute)
- INT16_2D — 2D 16-bit signed integer vectors (Short2Attribute)
- STRING — text strings (StringAttribute). String support is experimental and raises a warning when used: string attributes are accessible through the Python API but are not yet properly supported within Geometry Nodes. Values are read and written per-element (as unicode numpy arrays) since Blender’s foreach_get/foreach_set do not support string properties.
GeometrySet, for accessing all components of an object’s evaluated geometry — including multi-component Geometry Nodes outputs (mesh, point cloud, curves and instances) that a normally evaluated object doesn’t expose. Provides named_attribute() / list_attributes() across components and an informative repr() suited to snapshot testing.
Attribute.storage_type, exposing Blender 5.2’s new attribute storage types ("ARRAY" or "SINGLE"). SINGLE-storage attributes (produced by Geometry Nodes for constant values) read transparently as full arrays.
named_attribute(evaluate=True) and evaluate_object() now work for Curves and PointCloud objects, not just meshes.
BlenderObject.store_named_attribute() now returns the created or modified attribute (previously returned None).

Results of AttributeArray operations (e.g. pos + 1, comparisons) are now plain numpy.ndarrays, and copies (pos.copy(), boolean-mask indexing) are detached from Blender — neither syncs back on modification. True views (slices, column access) remain connected and continue to auto-sync.
Data is cast to the attribute’s storage dtype before writing, letting Blender’s foreach_set use its fast buffer path instead of per-element iteration (a significant speedup for the common case of passing float64 arrays).
Breaking: attribute type guessing for (n, 4) float arrays now maps to the generic FLOAT4 type instead of FLOAT_COLOR. To store colors or quaternions, explicitly request them with e.g. store_named_attribute(..., atype="FLOAT_COLOR"). (n, 4) uint8 arrays still map to BYTE_COLOR.
Attribute type guessing for (n, 2) int16 / uint16 arrays now maps to INT16_2D (other integer widths still map to INT32_2D), and string arrays map to STRING.
Requires bpy >= 5.2 (previously >= 5.1).
Improved type hints throughout the attribute module: precise Literal types for attribute type and domain names, tuple[int, ...] shapes, and an immutable AttributeType dataclass.

Setting an existing attribute with dictionary syntax (bob["color"] = data) no longer performs a second, redundant write with a re-guessed attribute type, which could raise a type mismatch error for color and quaternion attributes.
Attribute.from_array() now triggers the same object-data refresh workaround as store_named_attribute(), so writes through the low-level interface update reliably in the viewport.
repr() of an AttributeArray that has lost its attribute reference no longer raises an error.

Breaking: removed the unused legacy AttributeTypeInfo and AttributeDomain classes from the public API. Use the AttributeTypes and AttributeDomains enums instead.

Improved typing for the attribute module (#67).
Added move_to_collection() for moving objects from one collection into another (#62, thanks @kolibril13).

For older releases, see the GitHub releases page.