Lightning pose app

Welcome to the git repo for the lightning-pose-app.

This is a separate repo from core lightning-pose. You should not have to clone this unless changing source code.

📖 Documentation

Our full documentation, including installation guides, API references, and advanced tutorials, is available at:

👉 https://lightning-pose.readthedocs.io/

---

📝 Release Notes

[2.0.8.4] — 2026-04-20

Multiview EKS Model Support

This release introduces support for Ensemble Kalman Smoother (EKS) for multiview projects. It uses ensemble variance and spatiotemporal constraints to produce better predictions in cases where views disagree due to occlusions or overconfident tracking errors.

• Workflow:
  • Create an EKS model from the Models page. Select base models trained with different random seeds. This just creates a directory with the member models.
  • Run inference on all sessions to get predictions that can be viewed in the Viewer
• Seamless Integration: EKS model predictions follow the standard model directory schema.
• Chained Inference: Running inference on an EKS model automatically triggers inference for all constituent member models.

Limitations: Currently the app does not yet support non-linear EKS and singleview EKS. You can run these manually and view the predictions in the viewer as long as the model directory schema matches the schema expected by the app.

New "Run inference on all sessions" Flow:

• Replaces old inference that only ran inference on newly uploaded videos.
• Full coverage: Automatically processes all new sessions and models, skipping already-processed sessions, to ensure maximum model coverage in the Viewer.
• OOM Solution: Dynamically reduces batch size to prevent CUDA Out-of-Memory errors.
• Live Feedback: Displays real-time terminal output with an option to cancel.

Project management:

• Add existing project from the homepage
• Update existing project paths from the homepage
• New Directory picker UI replaces plain text path inputs

Other new features

• Global Task Indicator: Animated indicator in the navbar when there is active training or inference going on.
• Viewer > Upload session: Added ability to "Upload a new session" from the viewer
• Tensorboard Integration: Model details page now includes a "Tensorboard" tab with a ready-to-copy --logdir command.
• Viewer > videos all intra check: Added a warning when viewer detects videos that are not transcoded to the Viewer requirement of every-frame-is-intra-frame. This happens normally during video import but users may manually add videos to the directory and should be aware of it.

Bug Fixes

• Train/inference concurrency: Enforce a limit of one active inference or training session at a time.
• UI Status Polling: Fixed a bug where models remained in PENDING status; the list now correctly refreshes every 2.5s to show TRAINING.
• Path Persistence: Added automatic repair of stale data_dir and video_dir paths in model configs when a project is relocated.
• Video import normalization: Normalize Pixel Aspect Ratio to 1:1
• Video import singleview UX: no longer has leaks of multiview UI
• Video import session completeness validation: no longer allows importing sessions with some views missing

[2.0.8.3] — 2026-04-08

• Add viewer warning if video framerate or duration is inconsistent across views
• Fix 2.0.8.2 fps calculation
• Fix message in create model dialog about base

[2.0.8.2] — 2026-04-08

• Add ffprobe debug info to each video tile in the viewer.
• Fix 2.0.8.1 bug when multiview videos had different sizes

[2.0.8.1] — 2026-04-08

(Yanked, do not use)

• Fix viewer bug impacting accuracy of keypoint placement in rare scenarios.

[2.0.8.0] — 2026-04-07

• Redesigned Home page with a responsive project grid and rich project statistics (#65)
• Added project deletion: unregister vs. full delete (#65)
• Updated dropdown component: darker background, Popover API, fixed Actions button in models (#65)
• Added informational tooltips about file paths used in app: projects, labels, sessions (#73)
• Refined single-view labeler experience by hiding 'unknown' view (#73)
• Restored keypoint name tooltips in labeler and added keypoint label font size slider (#74)

[2.0.7.1] — 2026-03-17

• Refined labeler UX Fixed dragging off-center behavior, fixed subpixel rendering, and added keypoint text labels (#61)
• Added keypoint opacity and size controls to the labeler (#61)
• Added tooltips for data and model directories on project hover (#62)
• Added likelihood threshold control in viewer (#62)
• Improved display of calibration file information (#62)

[2.0.7.0] — 2026-02-26

• Added slider for video tile size in viewer
• Improved speed of loading label files and sessions lists by @hummuscience in #55
• Improved robustness of model creation date by @hummuscience in #55
• Fixed bug with saving empty labeling queue in #58

[2.0.5.4] — 2026-02-17

Added:

• Ability to resize left pane in Viewer, Labeler

Bugs Fixed:

• Labeled frames duplicated in unlabeled frames queue (#51)

[2.0.5.3] — 2026-02-10

Update (2026-02-17): This version introduced a bug #51 that was resolved in the next version, 2.0.5.4

Added:

• Extract frame from Viewer feature
  • You can now mark frames for labeling from the viewer, including the model's predictions on that frame.
• Remove frame from Label File or Label Queue feature
• Labeler View Options
  • See temporal context of frame (+/- 2 frames)
  • Image Brightness and Contrast slider

Behind the scenes

• Icons are now embedded in the app instead of CDN-hosted, so they can render locally without internet.
• Upgraded Angular from v20 to v21

[2.0.5.2] — 2026-02-04

Added

• Bundle Adjustment
  • Preview calibration before saving
  • Adjust extrinsics only by default, other algorithm improvements
• Model Management
  • Support for deleting and renaming models from the models table
  • Auto-open model logs after creation
• Misc
  • Added in-browser recommendation to use Chrome
  • Added release notes link in upgrade check

---

[2.0.5.1] — 2026-01-22

Added

• Extract frames: Manual frame selection. You can now specify a comma-separated list of frame indices to extract from a video.
• Support for uploading AVI files

Fixed

• Some usability quirks in the scroll-to-zoom viewports.

---

[2.0.5.0] — 2026-01-13

Initial Release