Changelog

The notable, user-facing changes for each release since v2.5.0. Bug fixes and
smaller internal tweaks are left out.

v2.6.0 — 2026-02-14

Full dual-channel LBM support — index, view, and save individual channels. The GUI is now channel-aware with a channel selector.
Read ImageJ hyperstacks directly as a TiffArray.
Dimension labels throughout, with time- and plane-based slicing and volumetric outputs.
Isoview metadata support, and OME-NGFF bumped to v0.5.

New IsoviewArray: a single 5D array covering raw, corrected, fused, and clusterpt data, read lazily one (t, c) at a time, and wired straight into the processing pipeline.
Internal 3D registration with Suite2p and GPU support.
Per-channel dimension and OME-NGFF compliance across the board.
Suite2p is now a required dependency instead of an optional one.
Isoview gains per-camera cropping at fuse time, projections written next to the tiles, streaming fusion to avoid out-of-memory crashes, and BigStitcher orientation control.

Non-destructive axial registration: apply or remove per-plane shifts at read time with with_axial_shifts, fully reversible.
Multi-resolution pyramids on write, with median and mode downsampling.
zscore support and a zarr I/O cleanup pass.
Dropped the default _stack suffix from output filenames.

Arrays are now always 5D (TCZYX). If you want the old trimmed shape, use arr.squeeze() or imread(path, squeeze=True).
.shape is the canonical 5D shape everywhere — the old .shape5d property has been removed.
Pluggable readers: third-party plugins can register their own formats and override the built-ins.
MP4Array unifies reading and writing mp4s; to_video now streams frame by frame.
GPU toggle via the MBO_GPU environment variable, the mbo gpu CLI command, or the Options panel. Warns if you select Cellpose without a GPU build of PyTorch.
Python 3.13 is now supported.

Full Changelog: v3.1.0...v3.2.0