Release Notes

0.1.13

For changes required in existing integrations, see Migrating from 0.1.12 to 0.1.13.

  • Added ConsensusWBFDetector with per-detector votes and optional weighted box fusion.

  • Preserved the winning case-sensitive spelling in majority-voting and medoid ensembles.

  • Added optional device-resident TrOCR CUDA decoding with explicit capability checks and NumPy fallback.

  • Reduced inference memory through batched crop preparation and detector preprocessing; added PPOCRRec width sorting and pixel budgets.

  • Added relative bbox padding and insets; enabled column inference in SimpleSorting by default.

  • Added optional native and browser runtimes, a browser OCR demonstration and explicit runtime installation extras.

  • Removed RapidFuzz and Levenshtein dependencies and refreshed training examples.

  • Replaced fixed blocks, lines and text spans with an extensible document tree using Element, Page and TextRegion.

  • Added table, cell and image elements with grid validation and orientation metadata.

  • Added single-image COCO serialization with Manuscript extensions and plain-text output from the tree.

  • Added Stack for composing processors and running independent detectors in parallel.

  • Added majority-voting and medoid ensembles with strict validation of recognizer output trees.

  • Added ReadingSorting for reading order and geometric containment with configurable overlap thresholds.

  • Updated SimpleSorting to group tree regions while retaining original objects and table contents.

  • Added configurable recognition levels, crop preparation and orientation compensation.

  • Added compatible PPOCRRec and TrOCR ONNX recognition, including TrOCR KV-cache decoding.

  • Added compatible PPOCR, RFDETR and Mask2Former detector integrations.

  • Added explicit detector class mappings and optional raw results, masks and original ONNX outputs.

  • Preserved native detection polygons and deprecated axis-aligned output options on EAST and YOLO.

  • Added a mirrored remote model registry with local caching, artifact verification and model-management helpers.

  • Added per-stage pipeline timings and retained independent intermediate Page snapshots.

  • Disabled correction by default and added explicit correction levels for CharLM.

  • Fixed partial ONNX batches and preserved configured recognizer batch sizes in Pipeline.

  • Improved Windows CUDA runtime loading and kept training dependencies out of ordinary inference imports.

  • Added PARSeq, SEResNet31LiteV2 and optional TPS rectification to TRBA training and export.

  • Improved checkpoint saving, training resume and final gradient-accumulation updates.

  • Updated tree visualization, grouping, text collapsing and image-input utilities.

  • Moved model export tools to repository scripts and updated the end-to-end demonstration interface.

  • Added notebooks for custom stages, model composition, training and export.

  • Expanded installation and test CI matrices through Python 3.14.

0.1.12

  • Added configurable TRBA batch sizes with a default of 128 and per-call overrides.

  • Updated Pipeline to respect the recognizer’s configured batch size.

  • Improved automatic CUDA DLL loading on Windows before ONNX session initialization.

  • Added create_page_from_image() for wrapping one or more images in a Page.

  • Added automatic ONNX export attempts after CharLM training.

  • Added onnxscript and Levenshtein to development dependencies and removed ultralytics.

0.1.11

  • Introduced a consistent Page -> Page stage contract and base classes and protocols for custom processors.

  • Refactored Pipeline around explicit stages and intermediate Page snapshots.

  • Added the YOLO ONNX detector with preset and custom weights and axis-aligned and rotated detections.

  • Set the default pipeline to YOLO with yolo26x_obb_text_g1, SimpleSorting and TRBA with trba_lite_g2.

  • Added the manuscript.layouts package and SimpleSorting, with configurable layout placement through layout_after.

  • Updated the data model around TextSpan, Line, Block and Page, retaining legacy Word aliases and versioned export schemas.

  • Added support for polygons with more than four points and helpers for merging and collapsing text spans.

  • Updated TRBA and CharLM to process Page objects with optional source-image context.

  • Added bbox, polygon_mask and quad_warp recognition crop presets and moved rotation and size filtering into TRBA.

  • Improved TRBA training, export, validation, metrics and debug tools, and added the trba_lite_g2 preset.

  • Improved EAST polygon parsing, training augmentations, previews, checkpoint selection and resume behavior.

  • Added HTTP model download retries with backoff.

  • Fixed Python 3.8 compatibility and training edge cases, including val_size=0 and pretrain_weights=None.

  • Updated API guides and example notebooks and introduced versioned documentation with archived 0.1.10 pages.

0.1.10

Reconstructed from published source distributions and Git history, relative to 0.1.8, the previous release on PyPI.

  • Added TRBA recognition and an end-to-end Pipeline combining detection, recognition and optional text correction.

  • Added the CharLM corrector with training, ONNX inference and export support.

  • Introduced structured OCR results using Word, Line, Block and Page, with JSON serialization.

  • Added shared model loading and caching for presets, local files, HTTP URLs, GitHub releases and Google Drive.

  • Reworked EAST inference around ONNX Runtime and added CPU, CUDA and CoreML provider selection for inference models.

  • Separated inference dependencies from optional PyTorch training and development dependencies.

  • Added configurable automatic rotation of vertical text crops before recognition.

  • Added optional EAST inference augmentation, contained-box filtering and area-anomaly filtering.

  • Improved EAST training resume and checkpoint recovery.

  • Added document sorting, reading-order visualization, OCR metrics and text-to-Page utilities.

  • Added a Gradio demonstration and expanded English and Russian API documentation.