SLEAP v1.6.5

SLEAP v1.6.5 is a small maintenance release: a critical fix restoring sleap-nn CLI functionality that had been silently disabled since v1.6.4, completion of the GUI's migration from the legacy track inference pipeline to the unified predict pipeline (fixing 3 related import bugs and a tracking-similarity dropdown flag drop along the way), sleap track now marked as a legacy command in favor of sleap predict, and removal of the redundant sleap-nn-* passthrough CLI entry points — plus an updated sleap-nn dependency (0.3.3) fixing a top-down inference scale-sharing regression plus two smaller correctness issues.

Quick install/upgrade:

uv tool install --python 3.13 "sleap[nn]==1.6.5" --torch-backend auto

See the v1.6.0 release notes for full details on the latest major release.

CLI Updates

Mark 'sleap track' as legacy, promote 'sleap predict' (#2839)

sleap track now prints a deprecation warning pointing users at sleap predict and is labeled (Legacy) in the CLI's help text. sleap predict is promoted throughout the docs as the recommended inference command, with its own full CLI reference section. track's underlying pipeline is unchanged — sleap-nn's own parity check found one confirmed behavioral divergence (tracking-ID counts on videos with empty-detection frames), so track isn't silently rerouted through predict; it's just clearly signposted as legacy so users can migrate at their own pace.

Remove sleap-nn-* passthrough entry points, warn on remaining legacy CLIs (#2842)

Removed the sleap-nn-train/sleap-nn-track/sleap-nn-export/sleap-nn-predict entry points, which duplicated the unified sleap CLI (sleap train/predict/export-model) without adding capability. Added a deprecation warning to the remaining legacy CLI commands that didn't already have one — sleap-render (→ sleap render), sleap-inspect (→ sleap show), and sleap-diagnostic (→ sleap doctor).

Bug Fixes

Fix sleap-nn CLI commands silently disabled by a stale import (#2836)

sleap/cli.py imported predict from sleap_nn.export.cli, but sleap-nn had moved predict into sleap_nn.cli. Because that import lived inside one broad try/except ImportError, the failure silently disabled all sleap-nn commands (train, track, eval, export-model, predict, system) — not just predict — so sleap.cli believed sleap-nn wasn't installed even when it was, and GUI training fell back to a stub command with Error: No such option '--config-name'.

Migrate GUI Run Inference to predict; fix 3 more broken imports and a tracking-similarity flag drop (#2838)

Completes the GUI's migration off sleap-nn's deprecated legacy track pipeline onto the unified predict pipeline (bit-for-bit equivalent per a full parity investigation). Along the way, fixes three more call sites broken by the same root cause as #2836 (sleap_nn.predict is a lazy top-level attribute, not an importable submodule) in sleap/nn_cli.py, sleap/legacy_cli_adaptors.py, and sleap/qc/insample_prediction.py. Also fixes a pre-existing bug where the tracking-similarity dropdown silently dropped --features/--scoring_method from the CLI call for any non-flow tracker using a similarity method other than IOU.

Dependency Updates

sleap-nn 0.3.2 → 0.3.3

Small correctness-focused follow-up, all fixes, no breaking changes:

• Fixed a regression where the top-down inference pipeline's centered-instance stage silently inherited the centroid stage's preprocessing.scale instead of using its own trained value, whenever the two models were trained at different scales (a common setup — centroid models are often trained at lower resolution for speed; this is literally the GUI's own default training-profile pairing). This corrupted confidence-map peak-finding badly enough to drop detections almost entirely: on a real 2560-frame project with mismatched scales, sleap-nn predict went from 5120 correct instances (v0.3.1) to 0 on v0.3.2. Both predict and track now correctly resolve and apply each stage's own scale by default, while still honoring an explicit --input_scale override applied uniformly to both stages (#725).
• Fixed sleap-nn predict never recording scale/crop_size in its output provenance metadata; for top-down models it now records centroid_scale/instance_scale/crop_size distinctly rather than collapsing to one shared value (#728).
• Fixed anchor_part having no upfront validation for centered_instance/multi_class_topdown/centered_instance_segmentation models — a typo'd or nonexistent anchor_part used to pass config setup cleanly and only fail deep inside dataset construction with a misleading error. A clear, correctly-attributed error is now raised upfront (#729).
• Fixed the LR scheduler selection silently ignoring its own documented priority order (cosine_annealing_warmup > linear_warmup_linear_decay > step_lr > reduce_lr_on_plateau) — any user who set a preferred scheduler without also explicitly nulling the always-populated-by-default reduce_lr_on_plateau silently got ReduceLROnPlateau instead, no error or warning. The scheduler is now selected in the documented priority order (#729).

See the sleap-nn v0.3.3 release notes for full details.

sleap-io — unchanged (0.9.2)

0.9.2 remains the latest sleap-io release; no bump needed this cycle.

Full Changelog: v1.6.4...v1.6.5

SLEAP v1.6.4

SLEAP v1.6.4 is a release packed with GUI improvements — Label QC label-error detection, a negative-frame training workflow, Merge Instance, motion-trail rendering controls, and per-instance visibility tools — alongside 25 bug fixes covering frame selection, skeleton persistence, drag-and-drop, Wayland/Windows dialog issues, and inference-pipeline crashes, plus updated dependencies (sleap-io 0.9.2 with re-identification, Category, and Event annotations; sleap-nn 0.3.1 with instance segmentation, sliding-window tiling, and pretrained backbones).

Quick install/upgrade:

uv tool install --python 3.13 "sleap[nn]==1.6.4" --torch-backend auto

See the v1.6.0 release notes for full details on the latest major release.

New Features

Add optional "Mean Node Score" column to Instances panel (#2696)

Adds an optional "Mean Node Score" column to the Instances panel, showing each predicted instance's mean point confidence over its visible nodes, mirroring sleap-nn's filter_min_mean_node_score filter. The column is hidden by default and toggled from View > Show Mean Node Score, giving users GUI visibility into a value previously only accessible via the CLI.

Improve loss plot y-axis scaling and add batch subsample dropdown (#2699)

Replaces a hard-coded floor in the training loss plot's log-scale y-axis calculation with log-space-aware padding so the loss curve fills the plot instead of being squeezed into a thin sliver, and adds a "Batch Subsample" dropdown (1/10/100) that thins out batch-loss scatter points for faster rendering on long training runs without discarding the underlying data.

Add "Include unlabeled frames" option to Render Video Clip dialog (#2701)

Adds an "Include unlabeled frames in range" checkbox to the Render Video Clip dialog, wiring the GUI through to sleap-io's existing full-range rendering semantics so exported clips can include every frame in a range (with overlays only where instances exist) instead of just labeled frames — previously only available via sio render --all-frames.

Add Accept All Predictions bulk action (#2702)

Adds an "Accept All Predictions..." action under the Labels menu that converts every unused predicted instance across the whole project into a user instance in one step, complementing the existing per-frame "Add Instances from All Predictions" command.

Expose peak threshold in inference dialog (#2704)

Exposes a Peak Threshold field (with a "Default" checkbox) in the GUI inference dialog, letting users tune the minimum confidence for node detections without dropping to the CLI's --peak_threshold flag.

Show current epoch runtime in training monitor (#2705)

The training monitor now shows per-epoch runtime alongside total runtime in the plot title, making it easy to spot a stalled epoch without manual arithmetic.

Add Previous User Labeled Frame navigation (Ctrl+Shift+U) (#2706)

Adds a "Previous User Labeled Frame" command that mirrors the existing "Next User Labeled Frame" (Ctrl+U), letting users navigate backward through frames containing user-created instances.

Add frame range filter for labeling suggestions (#2707)

Labeling suggestions can now be restricted to an explicit From/To frame range when the target is "Current Video," applying to every suggestion method except frame-chunk (which already has its own range control).

Add Actual Size (1:1) view option (#2708)

Adds a "View > Actual Size (1:1)" toggle that displays the video at native resolution instead of SLEAP's default zoom-to-fit, addressing complaints that low-resolution videos looked overly pixelated when upsampled to fit the window.

Add keyboard shortcut for Propagate Track Labels toggle (#2714)

Adds a default P keyboard shortcut for the "Propagate Track Labels" toggle, so users can flip that setting during proofreading without reaching for the Tracks menu each time.

Add Replace Videos button to Videos dock (#2715)

Adds a "Replace Videos" button to the Videos dock, alongside "Add Videos" and "Remove Video," so replacing a project's video no longer requires a trip to the File menu.

Negative frames — mark empty frames as background training examples (#2716)

Adds a full GUI workflow for marking a frame as a "negative frame" (confirmed-empty, no animals), including a toggle command, seekbar/canvas visual indicators, two new training-config options (Use Negative Frames, Negative Loss Weight), and a Label QC safety check. This closes a GUI gap for a capability sleap-io and sleap-nn already supported but that had been requested since 2022 (#640).

Label QC label-error detectors + GUI (#2770)

Adds a full Label QC label-error detection suite — five new detectors (left/right flip, chimera instances, duplicate/split instances, chain-order errors, missing nodes) — plus a revamped GUI with per-detector toggles, threshold sliders, filter-by-issue-type, and reviewed-state tracking. Also includes a spine-relative flip-detection fix that cut false-positive flip flags from 7.6% to 2.0% on real-world data.

Per-instance visibility & view-only toggles in the Instances dock (#2772)

Adds per-instance visibility and "view only" checkboxes to the Instances dock, letting users hide/show individual instances on the canvas or isolate a single instance for inspection without touching the saved project data.

Merge Instance — combine two partial instances into one (#2773)

Adds a "Merge Instance" action that combines two partially-labeled user instances in the same frame (e.g., one with anterior keypoints, one with posterior) into a single instance via shift/ctrl-select in the Instances dock, taking the union of their labeled nodes.

Per-instance & Label QC node-visibility controls (#2784)

Bundles three related Label GUI improvements on a shared per-instance node-visibility model: a Shift+V shortcut for "Show Non-Visible Nodes," a per-instance "Invisible Nodes" checkbox in the Instances dock, and a Label QC "Display" mode selector to reduce clutter when reviewing crowded frames.

Guide crop size / input scaling for top-down models (#2793)

Adds inline guidance to the training config dialog to steer users away from common top-down model mistakes — an info button explaining that input scaling should usually stay at 1.0 for centered-instance models, plus warnings when the effective crop size would fall below 100px or below the largest labeled instance. Also fixes a latent bug where validation warnings were rendered below the fold and effectively invisible.

Expose motion-trail options in the render/export dialog (#2819)

Adds a "Motion Trails" section to the Render Video Clip export dialog, exposing sleap-io's trail rendering options (length, node, width, opacity, fade, color) directly in the GUI so users no longer need the sio render --trails CLI to produce trailed videos.

Reframe live trail overlay onto sleap-io vocabulary (#2827)

Reframes the live in-player trail overlay to match sleap-io's trail semantics: trails now render for untracked/single-instance projects (previously required a Track and showed nothing), a new Trail Node menu lets users pick which node the trail follows, and fade is now a true alpha gradient. Replaces the removed "Trail Shade" preference with a "Trail Opacity" control.

CLI Updates

Show build commit hash in sleap doctor and startup banner (#2767)

sleap doctor and the startup banner now show the exact git commit a SLEAP install was built from (for editable and git-URL installs), and a new sleap doctor --commit flag resolves the commit for PyPI/conda releases via the GitHub API — useful for diagnosing installs that came from GitHub rather than a tagged release.

Bug Fixes

Fix Frame Selection and Frame Count in GUI (Shift + Drag and Shift + Double Click) (#2078)

Fixed Shift+Drag and Shift+Double-Click frame selection on the seekbar producing a frame count that didn't match the video's actual frame count, caused by an incorrect slider-width calculation and an extra +1 in the frame-count display.

Fix #2684: persist skeleton on save when no instances are labeled (#2686)

Fixed a skeleton edited only via the New Node/New Edge buttons being silently dropped on save when the project had no labeled instances yet, because the GUI's in-progress skeleton was never attached to labels.skeletons until an instance existed.

Fix blank progress dialog during Generate Suggestions (#2695)

Fixed "Generate Suggestions" blocking the main GUI thread and rendering its progress dialog as a black rectangle (Linux/Wayland) or triggering the spinning beach ball (macOS); suggestion generation now runs on a background thread with a responsive progress dialog.

Fix Copy Prior Frame copying stale prediction instead of user correction (#2698)

Fixed "Copy Prior Frame" landing on a stale, uncorrected prediction instead of the user's correction when both existed in the prior frame; the copy source now prefers user instances over unused predictions.

Fix parent widget on LearningDialog to fix window stacking on Windows (#2703)

Fixed training/inference dialogs appearing behind the main SLEAP window on Windows by giving LearningDialog a parent widget so Qt can manage window stacking correctly.

Clarify Delete Predictions beyond Frame Limit dialog text (#2711)

Clarified the "Delete Predictions beyond Frame Limit" dialog, whose title and field labels previously implied instances inside the range would be deleted — the opposite of the actual behavior.

Skip predicted instances in crop size auto-computation (#2717)

Crop size auto-computation in the training dialog now ignores PredictedInstance objects and accounts for augmentation padding, fixing cases where...

SLEAP v1.6.3

SLEAP v1.6.3 is a patch release with bug fixes for training launch, multi-video clip rendering, GUI exports for image-directory projects, prediction-to-user-label conversion, and updated dependencies (sleap-io 0.7.0 with a unified annotation architecture, sleap-nn 0.2.0 with negative-frame training and faster CLI startup).

Quick install/upgrade:

uv tool install --python 3.13 "sleap[nn]==1.6.3" --torch-backend auto

See the v1.6.0 release notes for full details on the latest major release.

Bug Fixes

Fix multi-GPU training from GUI on Windows and macOS (#2660)

Fixed multi-GPU training failing to launch from the GUI on Windows and macOS with ValueError: __main__.__spec__ is None. Training subprocesses are now invoked as a Python module (python -m sleap.cli train) instead of through the entry-point script, which lets PyTorch Lightning's DDP strategy correctly spawn worker processes.

Fix top-down training overwriting models when using a custom run name (#2659)

Fixed a bug where top-down training with a custom run_name saved both the centroid and centered-instance models into the same folder, causing the second model to overwrite the first and breaking inference. The model-type suffix (e.g., .centroid.n=42, .centered_instance.n=42) is now always appended so each model gets its own run folder.

Fix NaN-predicted nodes converting to visible user labels (#2676)

Fixed a bug where double-clicking a PredictedInstance (or running "Add instances from all predictions") could place nodes with NaN-predicted coordinates at random on-screen locations marked as visible, instead of as invisible user labels at a sensible fallback location. Two stacked bugs along the conversion path were fixed: uninitialized memory from Instance.empty() was leaking into "missing" node slots, and a parameter-shadowing bug in add_random_nodes could leak visible=True from a previously processed valid node onto subsequent NaN nodes.

Fix multi-video clip rendering: deduplication, frame order, source video, and FPS (#2671)

Fixed four independent bugs in the Render Video Clip dialog that combined to produce double skeletons, out-of-order frames, the wrong source video, and clips played at 1/4 real time on multi-video projects. User-corrected predictions are now deduplicated before rendering, frames are written in sorted order, the dialog's video selection is honored over the main-window selection, and a new "Match source video FPS" checkbox (on by default) keeps high-fps behavioral footage at real speed instead of clamping to 30 fps.

Fix crashes in export and replace-video dialogs for image-directory projects (#2669)

Fixed TypeError crashes in File > Export Analysis HDF5..., Export Analysis CSV..., and Replace Videos... when a project contained ImageVideo (image-directory) backends. These commands previously assumed video.filename was a single string, but ImageVideo returns a list of per-frame image paths; affected sites now normalize to a representative path before use.

Fix incorrect intensity augmentation defaults in training profiles (#2647)

Fixed legacy default values for intensity augmentation parameters (Gaussian noise, uniform noise, contrast, brightness) in all nine bundled training profiles, which would have produced unusable training images (fully black or washed-out) if the corresponding augmentations were enabled. Defaults are now standardized to a conservative range matching sleap-nn. Most users are unaffected because these augmentations are off by default, but anyone who had manually enabled them should now get sensible behavior.

Other

Use sleap-io for analysis HDF5 and CSV exports (#2649)

Refactored analysis HDF5 and CSV export (from the GUI and sleap-convert) to use sleap-io's save_analysis_h5() and save_csv() functions, replacing SLEAP's internal implementations. Output files remain backwards compatible (MATLAB axis ordering preserved), and exports from the SLEAP GUI now produce identical files to those from the sleap-io CLI.

Revise README badges and contributors section (#2661)

Removed the Conda downloads badge and reorganized the contributors section in the project README. Documentation/metadata change only; no user-facing functionality affected.

Dependency Updates

sleap-io 0.6.5 → 0.7.0

• Unified annotation architecture: Instance, BoundingBox, LabelImage, SegmentationMask, ROI, and Centroid are now first-class annotation types nested inside LabeledFrame, each with User* (ground-truth) and Predicted* (with score) variants and a uniform tracking_score field.
• First-class instance segmentation: New LabelImage type for Cellpose / StarDist / Mask R-CNN / SAM workflows, with streaming write, lazy read, multi-resolution scale/offset metadata, batch constructors (from_stack, from_binary_masks), and segmentation overlay rendering (API + sio render --overlay).
• First-class detection: New BoundingBox type with x1/y1/x2/y2 representation and full I/O across SLP, COCO, Ultralytics, GeoJSON, and JABS, plus rotated-box rendering.
• 3D pose data structures: New Identity (cross-session animal identity), Instance3D, and PredictedInstance3D for multi-camera workflows that round-trip with sleap-io.js and luc3d.
• New format support: Norpix .seq video files, TrackMate CSV reader (auto-detected by sio convert), h5wasm/sleap-io.js-written SLP files, and GeoJSON ROI I/O.
• Performance: O(1) frame and track index lookups across Labels, plus chunked SLP v2.2 label-image storage delivering ~43x faster writes and zero-decompression label-image merge.

See the sleap-io v0.7.0 release notes for full details.

sleap-nn 0.1.3 → 0.2.0

• sleap-io v0.7.0 adoption: Pinned to >=0.7.0,<0.8.0 to pick up the unified User*/Predicted* annotation architecture (drives the minor version bump).
• Negative frame training: Opt-in use_negative_frames lets user-confirmed empty frames suppress false-positive detections on SingleInstance, Centroid, BottomUp, and BottomUpMultiClass models, with per-sample loss metrics and cache-safe loading on Lustre/DDP/containers.
• Config picker parity and smarter defaults: sleap-nn config <slp> (TUI) now emits byte-identical YAML to the web-app config picker for all six pipelines, with corrected augmentation/head/trainer schema, matching max_stride recommendations, defaults flipped to Cache to Memory + 2 workers, and a simplified --pipeline topdown that emits paired centroid + centered_instance configs.
• sleap-nn info CLI: New subcommand prints a rich summary of any trained model directory — architecture, hyperparameters, training results, and evaluation metrics.
• Simpler install and faster CLI: torch/torchvision are now default dependencies (no [torch] extra needed; new [cpu]/[gpu] extras), and lazy imports cut sleap-nn -h startup from ~8s to ~1.2s.
• Training and inference fixes: ConvNeXt ONNX export no longer crashes on odd intermediate feature maps, multi-GPU training launched from GUIs is fixed via a __main__.__spec__ re-spawn check, check_memory() no longer reads every HDF5 frame (~21 min → <1 s), intensity augmentation defaults are back on the correct 0–1 scale, and node count / confidence filters now apply in track-only mode.
• Tracking robustness: hungarian_matching() no longer crashes on all-NaN cost matrices; non-finite values are replaced with large finite placeholders.
• Export and checkpoint metadata: ExportMetadata now carries an anchor_part field, and saved initial_config.yaml/training_config.yaml always reflect the running sleap_nn.__version__ at checkpoint time. TensorRT/ONNX export pipeline picked up additional bug fixes and broader test coverage.

See the sleap-nn v0.1.3 and v0.2.0 release notes for full details.

Full Changelog: v1.6.2...v1.6.3

SLEAP v1.6.2

SLEAP v1.6.2 is a patch release with bug fixes for GUI performance, data integrity, training configuration, and updated dependencies.

Quick install/upgrade:

uv tool install --python 3.13 "sleap[nn]==1.6.2" --torch-backend auto

See the v1.6.0 release notes for full details on the latest major release.

Bug Fixes

Fix GUI freeze when adding instances on suggested frames (#2632)

Fixed a performance regression that caused the GUI to freeze for ~10 minutes when adding instances on videos with large suggestion sets (~100k suggestions). The status bar update had O(n×m) complexity which has been reduced to O(n+m).

Fix skeleton node removal not updating instance point data (#2621)

Fixed a critical bug where deleting nodes from a skeleton via the GUI did not update instance point arrays. This caused file corruption where saved files couldn't be reopened due to shape mismatches. Files now correctly update point data when nodes are removed.

Improve training dialog UI with smart field visibility (#2619)

Training dialog improvements:

• Added toggle visibility for early stopping and OHKM parameter fields
• Hide OHKM fields for centroid models where they don't apply
• Fixed incorrect label "Sigma for Edges" → "Sigma for Identity" in multi-class bottom-up pipeline options

Fix Hydra override parsing error in exported train-script.sh (#2612)

Fixed OverrideParseException errors when running exported training scripts on SLURM clusters. The ckpt_dir and run_name values are now properly quoted to handle special characters.

Fix anchor part sync for top-down-id pipeline (#2610)

Fixed anchor_part selection not syncing correctly for the top-down-id pipeline in the training configuration dialog.

Dependency Updates

sleap-io 0.6.4 → 0.6.5

• ROI and Segmentation Mask support (experimental): New ROI class for vector geometry and SegmentationMask class for raster masks
• COCO Detection & Segmentation I/O (experimental): Read/write bounding box, polygon, and RLE mask annotations
• Ultralytics Detection & Segmentation I/O (experimental): Extended YOLO format support
• NumPy 2.x compatibility: Fixed serialization errors when saving .slp files

See the sleap-io v0.6.5 release notes for full details.

sleap-nn 0.1.0 → 0.1.2

• 6.7x faster bottom-up inference on NVIDIA A40 GPUs
• TUI Config Generator: Interactive wizard for generating training configs on remote systems
• Post-processing filters: New --filter_min_visible_nodes, --filter_min_mean_node_score options
• Multi-GPU fixes: Fixed DDP collective mismatch crashes and NCCL deadlocks
• Tracking fixes: Fixed track stealing bug and spurious track creation

See the sleap-nn v0.1.1 and v0.1.2 release notes for full details.

Full Changelog: v1.6.1...v1.6.2

SLEAP v1.6.1

SLEAP v1.6.1 is a patch release with bug fixes for Linux Qt compatibility, training configuration saving, and a new --video-backend CLI option.

Quick install/upgrade:

uv tool install --python 3.13 "sleap[nn]==1.6.1" --torch-backend auto

See the v1.6.0 release notes for full details on the latest major release.

Bug Fixes

Fix Linux Qt library conflicts (#2604)

On some Linux distributions (Debian 12, Fedora 43, and others with system Qt 6 packages), SLEAP could crash on launch with ImportError: undefined symbol errors due to conflicts between system Qt libraries and PySide6's bundled Qt. SLEAP now ensures PySide6's bundled Qt libraries take precedence on Linux by setting LD_LIBRARY_PATH and QT_PLUGIN_PATH before launching.

Fix training config save dialog bugs (#2603)

Fixed two bugs in the Training Configuration dialog reported in #2602:

• Run names were ignored when saving configs: User-entered run names were overwritten with auto-generated timestamps when using "Save configuration files..." or "Export training job package...". Run names are now preserved correctly.
• YAML file picker showed no files: The "Select training config file..." dropdown's file browser failed to show .yaml/.yml files due to an incorrect file filter separator. Fixed to use Qt's expected format.

New Features

--video-backend CLI option (#2604)

A new --video-backend flag allows selecting the video decoding backend when launching SLEAP:

sleap --video-backend FFMPEG

This is useful for working around h264 codec errors that can occur with OpenCV's default backend on some Linux systems. The setting persists across sessions via user preferences.

Other Changes

• Added workflow_dispatch trigger to the build workflow for manual CI re-runs
• Added display/GUI and video codec troubleshooting documentation

Full Changelog: v1.6.0...v1.6.1

What's New in SLEAP 1.6

SLEAP 1.6 is a major update with new backbone architectures, a redesigned training and inference experience, automated label quality control, ONNX/TensorRT export for faster deployment, a unified CLI, and MANY bug fixes. This release spans 70+ PRs since v1.5.2.

Quick start:

uv tool install --python 3.13 "sleap[nn]==1.6.0" --torch-backend auto

See below for more detailed installation instructions.

Major Changes

New Backbone Architectures in Training Dialog (#2579)

SLEAP 1.6 brings ConvNeXt and Swin Transformer backbone support to the GUI training dialog, alongside UNet:

• ConvNeXt -- Modern convolutional architecture in tiny/small/base/large variants (28M-198M parameters) with optional ImageNet pretrained weights for faster convergence.
• Swin Transformer (SwinT) -- Transformer-based backbone in tiny/small/base variants (28M-88M parameters) with optional ImageNet pretrained weights.

Select these from the training dialog's new backbone selector dropdown. See the sleap-nn model documentation for details.

Unified sleap CLI

SLEAP now has a single sleap command as the primary entry point. Running sleap launches the GUI, and subcommands provide access to all tools:

sleap              # Launch the GUI
sleap doctor       # System diagnostics and troubleshooting
sleap export-model # Export models to ONNX/TensorRT

All sleap-io and sleap-nn CLI commands are now integrated as sleap subcommands (#2524, #2541, #2559, #2587, #2595, #2597):

ONNX & TensorRT Model Export

Export trained models to optimized formats for 3-6x faster inference (#2573, #2594, #2595, #2597):

sleap export-model model.ckpt -o model.onnx --format onnx
sleap export-model model.ckpt -o model.engine --format tensorrt

Run inference on exported models:

sleap predict model.onnx video.mp4 -o predictions.slp

Benchmark results (NVIDIA RTX A6000, batch size 8):

To install export dependencies, reinstall with the appropriate extra: "sleap[nn,export]==1.6.0" (ONNX), "sleap[nn,export-gpu]==1.6.0" (ONNX + GPU), or "sleap[nn,tensorrt]==1.6.0" (TensorRT). See the sleap-nn Export Guide for full benchmarks and details.

Label Quality Control (#2547)

New sleap.qc module with GMM-based anomaly detection to automatically identify annotation errors. Accessible via Analyze > Label QC... in the GUI.

• Detects 10+ error types: isolated misses, jitter, visibility errors, scale issues, left-right swaps, gross misses, missing instances, and duplicates
• Dockable GUI widget with score histograms and sensitivity controls
• Keyboard navigation (Space/Shift+Space) to quickly review flagged instances
• Export to CSV or add flagged instances to Suggestions for batch review

Redesigned Training & Inference Dialogs (#2506, #2509, #2519, #2556, #2557, #2579)

The training and inference dialogs have been completely redesigned with native Qt for a faster, more polished experience:

• 55x faster config loading via rapidyaml and lazy loading
• Smaller dialog that fits on 1280x720 screens
• Augmentation controls simplified with on/off checkboxes and rotation presets (default: full ±180°)
• Device and worker settings default from user preferences
• Random Seed field for reproducible train/validation splits
• Evaluation metrics can be computed at configurable epoch intervals (mOKS, mAP, mAR, PCK, distance metrics logged to WandB)
• Prediction handling modes: Choose Keep, Replace, or Clear all predictions during inference
• "Random sample (current video)" inference target for quick model testing
• Form state persists after clicking Cancel

Real-Time Inference Progress (#2575)

The inference dialog now provides detailed progress feedback:

• Threaded inference: UI remains responsive during long-running jobs
• Live progress display: Predicted: 100/1,410 | FPS: 38.4 | ETA: 34s
• Log viewer: Dark-themed scrollable log showing subprocess output in real-time
• Working cancel button: Properly terminates the inference process
• "Delete All Predictions" now completes in milliseconds (was minutes on large datasets)

Video Rendering Overhaul (#2558)

Rendering is now powered by sleap-io's rendering engine with a live preview before exporting:

• 12+ new color palettes and 5 marker shapes
• Color by track, instance, or node
• Alpha transparency support for overlays
• Non-blocking video export with progress bar and cancel support

See the rendering documentation for details.

Filter Overlapping Instances (#2574)

New post-inference deduplication to remove duplicate predictions:

• IOU method: Filter by bounding box overlap
• OKS method: Filter by keypoint-based similarity
• Available in both the GUI (checkbox + threshold slider) and CLI (--filter_overlapping flags)

sleap doctor Diagnostics (#2524, #2551, #2553)

The sleap doctor command has been overhauled:

• Consolidated, copy-paste-friendly diagnostic output
• Git info display for editable installs (branch, commit hash)
• Comprehensive UV and conda introspection with conflict warnings
• System resources display (RAM and disk usage)
• -o/--output flag to save diagnostics to file
• Spinner during PyTorch import to show the command is working

How to Install

Step 1: Install uv (skip if already installed)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Install SLEAP v1.6.0

uv tool install --python 3.13 "sleap[nn]==1.6.0" --torch-backend auto

That's it! SLEAP is now available system-wide. The --torch-backend auto flag automatically detects your GPU (NVIDIA, AMD, Intel, or CPU). Run uv self update first if you get an error about this flag.

Step 3: Verify installation and launch

sleap doctor  # Check your setup
sleap          # Launch the GUI

Note: The sleap-label command from v1.5.x still works as an alias for launching the GUI.

Upgrading from v1.5.x?

uv tool install --reinstall --python 3.13 "sleap[nn]==1.6.0" --torch-backend auto

Quick data viewing (no permanent install)

uvx sleap labels.slp

Version compatibility

Note: Starting with SLEAP v1.5+, all deep learning functionality is powered by the PyTorch-based sleap-nn backend. TensorFlow models (with UNet backbones) from earlier versions are still supported for inference. Refer to the Migrating to 1.5+ docs for more details!

Dependency Updates

sleap-io v0.6.4 (#2597)

• ~90x faster SLP loading with lazy loading mode for large prediction files
• Pose rendering at ~50 FPS for publication-ready videos (sleap render)
• Data codecs for converting Labels to pandas DataFrames, NumPy arrays, and dictionaries
• Content-based video matching for reliable cross-platform merges even when file paths differ
• New Labels.match() API for inspecting matching results without merging
• Negative frames support: Mark frames as containing no instances (LabeledFrame.is_negative)
• 8 new CLI commands: merge, unsplit, fix, embed, unembed, trim, reencode, transform
• CSV format support for MATLAB interoperability
• Safe video matching prevents silent data corruption during merges
• Embedded images preserved during CLI operations (sio fix, sio convert, etc.)
• 23x faster .pkg.slp saves, 2.7x faster embedded video loading
• Bug fixes for video matching, rendering, embedded videos, skeleton consolidation

sleap-nn v0.1.0 (#2597)

• ONNX/TensorRT export for 3-6x faster inference
• Epoch-end evaluation metrics: mOKS, mAP, mAR, PCK, and distance metrics logged to WandB
• Post-inference filtering: Greedy NMS to remove duplicate predictions
• 17-51x faster peak refinement (integral refinement now works on Mac)
• GUI progress mode: JSON output for real-time progress in SLEAP GUI
• Faster inference via GPU-accelerated normalization (17% typ...

SLEAP v1.6.0a3

About the v1.6 Pre-release Series

This is a pre-release for SLEAP v1.6.0. It contains many new features and improvements, but is not yet considered stable. For production use, see v1.5.2.

We are releasing a series of pre-releases that incrementally build towards the stable v1.6.0 release. Each pre-release adds new features and bug fixes:

Note: Starting with SLEAP v1.5+, all deep learning functionality is powered by the PyTorch-based sleap-nn backend. TensorFlow models (with UNet backbones) from earlier versions are still supported for inference. Refer to the Migrating to 1.5+ docs for more details!

How to Install

Step 1: Install uv (skip if already installed)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Install SLEAP v1.6.0a3

uv tool install --python 3.13 "sleap[nn]==1.6.0a3" --with "sleap-io==0.6.3" --with "sleap-nn==0.1.0a4" --prerelease allow --torch-backend auto

That's it! SLEAP is now available system-wide. The --torch-backend auto flag automatically detects your GPU (NVIDIA, AMD, Intel, or CPU). Be sure to do a uv self update if you get an error about this flag.

Step 3: Verify installation

sleap doctor

Upgrading from v1.6.0a2?

uv tool upgrade sleap --upgrade-package sleap-io --upgrade-package sleap-nn

Or for a clean reinstall:

uv tool install --reinstall --python 3.13 "sleap[nn]==1.6.0a3" --with "sleap-io==0.6.3" --with "sleap-nn==0.1.0a4" --prerelease allow --torch-backend auto

Rollback to stable

If you encounter issues, rollback to the latest stable release:

uv tool install --python 3.13 "sleap[nn]==1.5.2" --torch-backend auto

Version compatibility

What's New in v1.6.0a3

New Backbone Architectures (#2579)

Train models using modern transformer-based and ConvNeXt architectures as alternatives to UNet:

• ConvNeXt: Select from tiny/small/base/large variants (28M-198M parameters) with optional ImageNet pretrained weights for faster convergence
• Swin Transformer (SwinT): Select from tiny/small/base variants (28M-88M parameters) with optional ImageNet pretrained weights

Access these from the training dialog's new backbone selector dropdown.

ONNX/TensorRT Model Export (#2573)

Export trained models to optimized formats for 3-6x faster inference:

sleap-nn-export model.ckpt -o model.onnx --format onnx
sleap-nn-export model.ckpt -o model.engine --format tensorrt

Run inference on exported models:

sleap-nn-predict model.onnx video.mp4 -o predictions.slp

Benchmark results (NVIDIA RTX A6000, batch size 8):

See the sleap-nn Export Guide for full benchmarks and usage details.

Real-Time Inference Progress (#2575)

The inference dialog now provides detailed progress feedback:

• Threaded inference: UI remains responsive during long-running jobs
• Live progress display: Predicted: 100/1,410 | FPS: 38.4 | ETA: 34s
• Log viewer: Dark-themed scrollable log showing subprocess output in real-time
• Working cancel button: Properly terminates inference when clicked

Filter Overlapping Instances (#2574)

New GUI controls to remove duplicate/overlapping predictions after inference:

• Enable filtering checkbox in the Preprocessing/Postprocessing section
• Method selection: IOU (bounding box overlap) or OKS (keypoint-based similarity)
• Threshold control: Lower values = more aggressive filtering (default: 0.8)

Also available via CLI:

sleap-nn-track model.ckpt video.mp4 --filter_overlapping --filter_overlapping_method oks --filter_overlapping_threshold 0.5

Evaluation During Training (#2579)

New evaluation section in the training dialog:

• Enable evaluation checkbox to compute metrics during training
• Frequency control to set how often evaluation runs (in epochs)
• Metrics logged to WandB: mOKS, mAP, mAR, PCK, distance percentiles

Performance Improvements

• Delete All Predictions (#2575): Now completes in milliseconds instead of minutes on large datasets
• 17-51x faster peak refinement in sleap-nn (v0.1.0a4): Enables integral refinement on Mac (previously disabled)

Bug Fixes

macOS Fixes

• Fixed crash when opening training dialog on macOS with Homebrew installed. The crash was caused by a conflict between Homebrew's libpng and macOS's ImageIO framework during Qt font rendering. (#2571)
• Fixed dialog button ordering on macOS. Training and inference dialog buttons now appear in consistent order across all platforms (Mac, Windows, Linux). (#2576)
• Fixed default button highlighting: The "Run" button now correctly appears as the default (highlighted) button instead of "Copy to clipboard". (#2576)

UI Fixes

• Fixed dark mode for training dialog main tab. The background now properly follows the system theme instead of remaining white. (#2572)
• Fixed loss monitor to recognize sleap-nn's metric naming convention (train/loss, val/loss). (#2579)

Other Fixes

• Fixed ConvNeXt/SwinT training crash in sleap-nn: Resolved skip connection channel mismatch that caused errors during validation. (sleap-nn v0.1.0a4)
• Fixed inference progress ending at 99%: Now correctly shows 100% when complete. (sleap-nn v0.1.0a4)
• Fixed CSV learning rate logging: The learning_rate column in training_log.csv is no longer empty. (sleap-nn v0.1.0a4)

Dependency Updates

sleap-nn v0.1.0a4

Changes since v0.1.0a2 (the previous minimum version):

• ONNX/TensorRT Export: Export models to optimized formats for 3-6x faster inference
• Post-Inference Filtering: Greedy NMS to remove duplicate predictions (--filter_overlapping)
• 17-51x Faster Peak Refinement: Fast tensor indexing replaces kornia's crop_and_resize
• GUI Progress Mode: New --gui flag enables JSON output for real-time GUI progress
• Simplified Train CLI: sleap-nn train config.yaml (positional config path)
• Bug fixes for ConvNeXt/SwinT training, CSV logging, progress display

sleap-io v0.6.3

Changes since v0.6.2 (the previous minimum version):

• Negative Frames Support: Mark frames as containing no instances (LabeledFrame.is_negative)
• Embedded Images Preserved: CLI commands (sio fix, sio convert, etc.) no longer strip embedded images
• Smart Skeleton Consolidation: Compatible skeletons are reassigned instead of deleted

Full Changelog

Enhancements

• Add ConvNeXt and SwinT backbone options to training dialog by @talmo in #2579
• Improve inference dialog with real-time progress and UI fixes by @talmo in #2575
• Add filter_overlapping controls to training/inference dialogs by @talmo in #2574
• Bump sleap-io to 0.6.3 and sleap-nn to 0.1.0a3, add new CLI commands by @talmo in #2573

Fixes

• Fix macOS crash caused by Homebrew libpng conflict by @talmo in #2571
• Fix training dialog main tab background to match config tabs by @talmo in #2572
• Fix Mac dialog button order and default button styling by @talmo in #2576

Workflows

• A...

SLEAP v1.6.0a2

About the v1.6 Pre-release Series

This is a pre-release for SLEAP v1.6.0. It contains many new features and improvements, but is not yet considered stable. For production use, see v1.5.2.

We are releasing a series of pre-releases that incrementally build towards the stable v1.6.0 release. Each pre-release adds new features and bug fixes:

Note: Starting with SLEAP v1.5+, all deep learning functionality is powered by the PyTorch-based sleap-nn backend. TensorFlow models (with UNet backbones) from earlier versions are still supported for inference. Refer to the Migrating to 1.5+ docs for more details!

How to Install

Step 1: Install uv (skip if already installed)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Install SLEAP v1.6.0a2

uv tool install --python 3.13 "sleap[nn]==1.6.0a2" --with "sleap-io==0.6.2" --with "sleap-nn==0.1.0a2" --prerelease allow --torch-backend auto

That's it! SLEAP is now available system-wide. The --torch-backend auto flag automatically detects your GPU (NVIDIA, AMD, Intel, or CPU). Be sure to do a uv self update if you get an error about this flag.

Step 3: Verify installation

sleap doctor

Upgrading from v1.6.0a1?

uv tool upgrade sleap --upgrade-package sleap-io --upgrade-package sleap-nn

Or for a clean reinstall:

uv tool install --reinstall --python 3.13 "sleap[nn]==1.6.0a2" --with "sleap-io==0.6.2" --with "sleap-nn==0.1.0a2" --prerelease allow --torch-backend auto

Rollback to stable

If you encounter issues, rollback to the latest stable release:

uv tool install --python 3.13 "sleap[nn]==1.5.2" --torch-backend auto

Version compatibility

What's New in v1.6.0a2

• Revamped Installation Documentation:

  • Complete rewrite of installation docs with simplified workflow (#2567)
  • Single universal install command for all platforms using --torch-backend auto
  • Reduced from 8 installation paths to 2 (tool install + development setup)
  • New uvx sleap labels.slp option for viewing data without permanent installation
  • Streamlined upgrade flow with uv tool upgrade sleap

• Python 3.13 Default:

  • Python 3.13 is now the default recommended version (#2565)
  • Python 3.12 remains supported

• Bug Fixes:

  • Fixed "Export Training Job Package" crash with ConfigAttributeError: Missing key zmq that occurred on v1.6.0a0/v1.6.0a1 (#2566, fixes #2562)

• sleap-io v0.6.2:

  • Content-Based Video Matching: Videos are now automatically matched by pose annotations or pixel content, enabling reliable cross-platform merges even when file paths differ
  • New Labels.match() API: Inspect matching results without merging — ideal for evaluation workflows
  • Video Color Mode Control: New Labels.set_video_color_mode() method and sio fix --video-color CLI option
  • Bug fixes for HDF5 dataset matching and provenance conflict handling

• sleap-nn v0.1.0a2:

  • Epoch-End Evaluation Metrics: Real-time mOKS, mAP, mAR, PCK, and distance metrics logged to WandB during training
  • Robust Video Matching: Uses sleap-io's Labels.match() API for better cross-platform evaluation
  • Bug fixes for embedded video handling and centroid model ground truth matching

Full Changelog

Enhancements

• Revamp installation documentation for uv-based workflow by @talmo in #2567

Fixes

• Fix export training package zmq config error by @talmo in #2566

Workflows

• Fix docs workflow race condition with concurrency group by @talmo in #2563
• Housekeeping: Python 3.13 default and sleap-support skill by @talmo in #2565

Dependencies

• Bump sleap-io min to 0.6.2 and sleap-nn min to 0.1.0a2 by @talmo in #2569

Full Changelog: v1.6.0a1...v1.6.0a2

SLEAP v1.6.0a1

About the v1.6 Pre-release Series

This is a pre-release for SLEAP v1.6.0. It contains many new features and improvements, but is not yet considered stable. For production use, see v1.5.2.

We are releasing a series of pre-releases that incrementally build towards the stable v1.6.0 release. Each pre-release adds new features and bug fixes:

Note: Starting with SLEAP v1.5+, all deep learning functionality is powered by the PyTorch-based sleap-nn backend. TensorFlow models (with UNet backbones) from earlier versions are still supported for inference. Refer to the Migrating to 1.5+ docs for more details!

How to Install

Step 1: Install uv (skip if already installed)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Install SLEAP v1.6.0a1

Windows/Linux with NVIDIA GPU (CUDA 12.8)

uv tool install --reinstall --python 3.12 "sleap[nn]==1.6.0a1" --with "sleap-io==0.6.1" --with "sleap-nn==0.1.0a1" --prerelease allow --index https://download.pytorch.org/whl/cu128 --index https://pypi.org/simple

Windows/Linux with NVIDIA GPU (CUDA 13.0)

uv tool install --reinstall --python 3.12 "sleap[nn]==1.6.0a1" --with "sleap-io==0.6.1" --with "sleap-nn==0.1.0a1" --prerelease allow --index https://download.pytorch.org/whl/cu130 --index https://pypi.org/simple

Windows/Linux without GPU (CPU only)

uv tool install --reinstall --python 3.12 "sleap[nn]==1.6.0a1" --with "sleap-io==0.6.1" --with "sleap-nn==0.1.0a1" --prerelease allow --index https://download.pytorch.org/whl/cpu --index https://pypi.org/simple

macOS

uv tool install --reinstall --python 3.12 "sleap[nn]==1.6.0a1" --with "sleap-io==0.6.1" --with "sleap-nn==0.1.0a1" --prerelease allow

Step 3: Verify installation

sleap doctor

Upgrading from v1.6.0a0?

Use the same commands as above. The --reinstall flag will create a clean environment with the new dependencies.

Rollback to stable

If you encounter issues, rollback to the latest stable release:

# Windows/Linux (CUDA 12.8)
uv tool install --reinstall --python 3.12 "sleap[nn]==1.5.2" --index https://download.pytorch.org/whl/cu128 --index https://pypi.org/simple

# Windows/Linux (CPU only)
uv tool install --reinstall --python 3.12 "sleap[nn]==1.5.2" --index https://download.pytorch.org/whl/cpu --index https://pypi.org/simple

# macOS
uv tool install --reinstall --python 3.12 "sleap[nn]==1.5.2"

Version compatibility

What's New in v1.6.0a1

• Label Quality Control (QC):

  • New sleap.qc module with GMM-based anomaly detection to automatically identify annotation errors (#2547)
  • Detects 10+ error types: isolated misses, jitter, visibility errors, scale issues, left-right swaps, gross misses, missing instances, and duplicates
  • Dockable GUI widget accessible via Analyze > Label QC... with score histograms and sensitivity controls
  • Keyboard navigation (Space/Shift+Space) to quickly navigate flagged instances
  • Export to CSV or add flagged instances to Suggestions for review

• Enhanced CLI:

  • 8 new CLI commands from sleap-io: sleap merge, sleap unsplit, sleap fix, sleap embed, sleap unembed, sleap trim, sleap reencode, sleap transform (#2559)
  • See the sleap-io CLI documentation for detailed usage

• Video Rendering Overhaul:

  • Now powered by sleap-io's rendering engine — see rendering documentation for details (#2558)
  • Live preview of rendered frames with all style options before exporting
  • 12+ new color palettes and 5 marker shapes with options to color by track, instance, or node
  • Alpha transparency support for overlays
  • Non-blocking video export with progress bar and cancel support

• Training Dialog Improvements:

  • Form state now persists after clicking Cancel (#2557)
  • Device and worker settings default from user preferences instead of being overwritten by profiles (#2557)
  • Updated all baseline profiles to use full ±180° rotation augmentation (#2557)
  • Added Random Seed field for reproducible train/validation splits (#2557)
  • New tooltips for Input Scaling, Batch Size, Predict On, and tracker fields (#2556)

• Inference Improvements:

  • New "Random sample (current video)" inference target for quick model testing (#2555)
  • Fixed --exclude_user_labeled flag not working when "Skip user labeled frames" is enabled (#2552)

• sleap doctor Improvements:

  • Consolidated, copy-paste-friendly diagnostic output (#2553)
  • Git info display for editable installs (branch, commit hash) (#2553)
  • Comprehensive UV and conda introspection with conflict warnings (#2553)
  • System resources display (RAM and disk usage) (#2553)
  • New -o/--output flag to save diagnostics to file (#2553)
  • Added spinner during PyTorch import to show command is working (#2551)
  • Fixed path truncation in tables (#2551)

• Bug Fixes:

  • Fixed terminal spam from "Error processing frame" messages when scrubbing .pkg.slp files (#2554)

• sleap-io v0.6.1:

  • 8 new CLI commands: merge, unsplit, fix, embed, unembed, trim, reencode, transform
  • CSV format support for MATLAB interoperability
  • Coordinate-aware video transformations
  • 23x faster .pkg.slp saves, 2.7x faster embedded video loading
  • Bug fixes for video matching, rendering, and embedded videos

• sleap-nn v0.1.0a1:

  • Training progress bar during dataset caching (no more apparent "freeze")
  • Automatic WandB local log cleanup to save disk space
  • Simplified log format for cleaner output

Full Changelog

Enhancements

• Add sleap.qc module for label quality control by @talmo in #2547
• Add sleap-io v0.6.1 CLI commands by @talmo in #2559
• Upgrade video rendering to use sleap-io API with live preview and non-blocking progress by @talmo in #2558
• Improve training config dialog UX by @talmo in #2557
• Add missing tooltips to training config and tracker form fields by @talmo in #2556
• Add "Random sample (current video)" inference target option by @talmo in #2555
• Improve sleap doctor with consolidated diagnostic output by @talmo in #2553
• Improve sleap doctor UX: add spinner and fix path truncation by @talmo in #2551

Fixes

• Suppress frame error spam when scrubbing pkg.slp files by @talmo in #2554
• Add --exclude_user_labeled flag to sleap-nn-track CLI shim by @talmo in #2552

Workflows

• Add PR-local docs preview deployment by @talmo in #2550

Dependencies

• Bump version to 1.6.0a1 by @talmo in #2561

Full Changelog: v1.6.0a0...v1.6.0a1

SLEAP v1.6.0a0

About the v1.6 Pre-release Series

This is a pre-release for SLEAP v1.6.0. It contains many new features and improvements, but is not yet considered stable. For production use, see v1.5.2.

We are releasing a series of pre-releases that incrementally build towards the stable v1.6.0 release. Each pre-release adds new features and bug fixes:

Note: Starting with SLEAP v1.5+, all deep learning functionality is powered by the PyTorch-based sleap-nn backend. TensorFlow models (with UNet backbones) from earlier versions are still supported for inference. Refer to the Migrating to 1.5+ docs for more details!

How to Install

Step 1: Install uv (skip if already installed)

# Windows
powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS/Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Step 2: Install SLEAP v1.6.0a0

Windows/Linux with NVIDIA GPU (CUDA 12.8)

uv tool install --force --python 3.12 "sleap[nn]==1.6.0a0" --with "sleap-io==0.6.0" --with "sleap-nn==0.1.0a0" --prerelease allow --index https://download.pytorch.org/whl/cu128 --index https://pypi.org/simple

Windows/Linux with NVIDIA GPU (CUDA 13.0 - NEW!)

uv tool install --force --python 3.12 "sleap[nn]==1.6.0a0" --with "sleap-io==0.6.0" --with "sleap-nn==0.1.0a0" --prerelease allow --index https://download.pytorch.org/whl/cu130 --index https://pypi.org/simple

Windows/Linux without GPU (CPU only)

uv tool install --force --python 3.12 "sleap[nn]==1.6.0a0" --with "sleap-io==0.6.0" --with "sleap-nn==0.1.0a0" --prerelease allow --index https://download.pytorch.org/whl/cpu --index https://pypi.org/simple

macOS

uv tool install --force --python 3.12 "sleap[nn]==1.6.0a0" --with "sleap-io==0.6.0" --with "sleap-nn==0.1.0a0" --prerelease allow

Step 3: Verify installation

sleap doctor

Upgrading from v1.5.x?

Use the same commands as above. The --force flag will replace your existing installation.

Rollback to stable

If you encounter issues, rollback to the latest stable release:

# Windows/Linux (CUDA 12.8)
uv tool install --force --python 3.12 "sleap[nn]==1.5.2" --index https://download.pytorch.org/whl/cu128 --index https://pypi.org/simple

# Windows/Linux (CPU only)
uv tool install --force --python 3.12 "sleap[nn]==1.5.2" --index https://download.pytorch.org/whl/cpu --index https://pypi.org/simple

# macOS
uv tool install --force --python 3.12 "sleap[nn]==1.5.2"

Version compatibility

What's New in v1.6.0a0

• Unified CLI:

  • New sleap command as the primary entry point - just run sleap to launch the GUI (#2524)
  • Integrated sleap-io CLI commands: sleap show, sleap convert, sleap split, sleap filenames, sleap render (#2541)
  • New sleap doctor command for system diagnostics and troubleshooting (#2524)

• Training GUI Overhaul:

  • 55x faster config loading and much faster dialog startup (#2506, #2516)
  • Completely redesigned training dialog with native Qt, unified frame targeting, and 356 new tests (#2519)
  • New Frame Target Selector for flexible training/inference frame selection (#2519)
  • Prediction handling modes: Keep, Replace, or Clear all predictions during inference (#2519)
  • Smaller dialog that fits on 1280x720 screens (#2509, #2519)
  • Augmentation controls simplified with on/off checkboxes and rotation presets (#2509)
  • WandB integration improvements with run URL display and auto-browser-open (#2525)

• Crop Size Visualization:

  • Visual crop size preview in training dialog for top-down models (#2483)
  • New Instance Size Distribution widget for analyzing bounding box sizes across your dataset (#2528)
  • Click-to-navigate from size distribution plot to specific instances (#2528)

• New Features:

  • "Check for Updates" dialog showing versions for sleap, sleap-io, and sleap-nn (#2499)
  • "Delete Predictions on User-Labeled Frames" for cleaning up duplicate instances (#2505)
  • Startup banner with version info and branding when launching sleap-label (#2517)
  • Progress dialog with cancel support for package export (#2522)
  • Support for loading legacy SLEAP metrics from v1.4.1 and earlier (#2480)

• Critical Bug Fixes:

  • Fixed GUI freeze when editing predictions with NaN coordinates on Linux Qt 6.10+ (#2467)
  • Fixed catastrophic data loss bug where removing a video could delete frames from ALL videos with the same resolution (#2535)
  • Fixed prediction deletion incorrectly removing user-labeled instances (#2478)
  • Fixed predictions not being fully converted to instances when adding from predictions (#2539)

• sleap-io v0.6.0: (#2540)

  • ~90x faster SLP loading with new lazy loading mode for large prediction files
  • Pose rendering at ~50 FPS for publication-ready videos (sleap render)
  • Data codecs for converting Labels to pandas DataFrames, NumPy arrays, and dictionaries
  • Safe video matching prevents silent data corruption during merges
  • Fixed package export losing videos without labeled frames (#282)
  • Fixed video provenance breaking during merge operations (#302)
  • Breaking: Merge API simplified (video_matcher= → video=, frame_strategy= → frame=)

• sleap-nn v0.1.0a0: (#2545)

  • Faster inference via GPU-accelerated normalization (17% for typical video, up to 50% for large RGB images)
  • CUDA 13.0 support for latest NVIDIA GPUs
  • Provenance tracking embeds full reproducibility metadata in output files
  • Enhanced WandB with interactive visualizations and per-head loss logging
  • Fixed crash on frames with empty instances (#385)
  • Fixed --exclude_user_labeled being ignored with --video_index (#397)
  • Fixed run folder cleanup when training canceled via GUI (#392)
  • Breaking: Crop size semantics changed - top-down models now crop first, then resize
  • Breaking: Output file naming changed (labels_train_gt_0.slp → labels_gt.train.0.slp)

• Other dependency changes:

  • Removed 8 unused dependencies for faster installation (#2486)

Full Changelog

Enhancements

• Add unified CLI with sleap command by @talmo in #2524
• Add sleap-io CLI command inheritance (show, convert, split, filenames, render) by @talmo in #2541
• Add "Check for Updates" to Help menu and implement update checker by @jaw039 in #2499
• Refactor training/inference dialog with native Qt and unified frame targeting by @talmo in #2519
• Training GUI QOL improvements (augmentation checkboxes, rotation presets, overfit mode) by @talmo in #2509
• Improve training dialog startup performance (~55x faster config loading) by @talmo in #2506
• Add crop size visualization for top-down training pipelines by @gitttt-1234 in #2483
• Add Instance Size Distribution widget for crop size analysis by @talmo in #2528
• Add Delete Predictions on User-Labeled Frames feature by @talmo in #2505
• Add support for loading legacy SLEAP metrics (<=v1.4.1) by @gitttt-1234 in #2480
• Add progress dialog and completion notification for package export by @talmo in #2522
• Add startup banner with version info and branding by @talmo in #2517
• Improve baseline config display names in training config selector by @gitttt-1234 in #2471
• Fix WandB checkbox state and add run URL display by @talmo in #2525

Fixes

• Fix GUI freeze when editing predictions with NaN coordinates by @gitttt-1234 in #2467
• Fix remove_video() to use identity comparison instead of matches_content() by @talmo in #2535
• Fix prediction deletion to prevent removing labeled instances by @gitttt-1234 in #2478
• Add failing tests for predictions-not-fully-added bugs (and fix) by @talmo in #2539
• Fix GUI freeze during labeled video export by @gitttt-1234 in #2484
• Fix skeleton loading returning list instead of single Skeleton by @gitttt-1234 in #2493
• Fix missing file dialog for ImageVideo backend (list of frame paths) by @alicup29 in #2498
• Fix delete unused tracks crash with untracked instances by @talmo in #2503
• Fix plateau detection to use absolute threshold mode by @gitttt-1234 in #2469
• Update predictions output path for inference (multi-v...