Version History

Development timeline and changes

Latest Release

Version 0.1.4 (October 2026)

Process OCR

  • Auto-fix OCR typos: when the recognition ends, the labels of the project are compared with each other and look-alike misreadings are corrected by the most frequent spelling (1NV. or INU → INV., POB → PDB, VS → US, 5.4 → S.4). Numbers are never changed, and the labels you corrected by hand are never rewritten
  • Fixed labels show the corrected text with the changed parts highlighted, the original reading below, and an Undo button. The switch below the toolbar turns the step off; Fix OCR Typos runs it on results you already have
  • The raw OCR is kept: the fixes go into the corrected text, so the catalogue still has both versions

Review Texts and Export

  • Labels fixed automatically carry an Auto-corrected badge; editing one by hand turns it into a manual correction
  • The Manual Corrections counter no longer counts the automatic fixes

Version History

Version 0.1.3 (October 2026)

Annotate

  • Auto-detect: a YOLO detector (about 19 MB, downloaded at startup) finds the drawings and the labels on a plate and attaches each label to its nearest drawing, ready for you to check
  • Right-click menus on the canvas: assign a text to another drawing, delete a text, add a text to a drawing or delete it. 1-9 reassigns the selected text box
  • Boxes follow the cursor outside the canvas and stop at the image edge; the Delete Box button no longer wraps the toolbar
  • Opening or deleting a project now resets all of the previous project’s state

Load, OCR and Clean

  • Large folders are uploaded in batches, which fixes the 413 error
  • The OCR model can be switched from the Projects page after the first launch; GLM-OCR no longer outputs non-Latin or LaTeX tokens
  • Clean Drawings is faster with many drawings: thumbnails load lazily, undo copies the canvas, and a drawing is saved only if you edited it. Review Texts loads its previews lazily

Version 0.1.2 (September 2026)

OCR and models

  • GLM-OCR: a lightweight local OCR model (0.9B parameters) that runs on CPU or GPU, so an NVIDIA GPU is no longer required. It is the recommended choice on first launch; OlmOCR-FP4 remains available for NVIDIA GPUs, and you can still start without any OCR model
  • New setup screen with a progress bar per model and recovery actions when a download fails: Retry Download, Choose a Different Model, Continue Without OCR
  • The few-shot parsing model is now Qwen3.5-2B (it was Qwen3-1.7B). If its download fails the app still starts, and only the Parser Studio is unavailable
  • Process OCR tab redesigned: live status while the boxes are read, the recognized labels listed as they arrive, a log drawer, Copy Extracted Texts and Re-run OCR Processing

Annotate and Load

  • Every change is saved automatically, with a Saved badge; a mode badge shows which tool is active
  • Move a selected box with the arrow keys (2 px, or 10 px with Shift), and delete it with the Delete Box button or Delete
  • Drag and drop of files and folders in the Load tab, a Plate Catalog of thumbnails, and Add Files / Add Folder to extend a project at any time
  • In-app confirmation dialogs replace the browser pop-ups

Clean Drawings

  • Straighten tool: precision slider (−45° to +45°), micro-adjustment buttons, quarter-turn rotations, an alignment grid and Auto-Fit to trim the blank margins
  • Eraser presets (6, 16, 32 and 64 px), a brush preview circle, mouse-wheel size and zoom, and new shortcuts (G, R, [, ], +, -)

Export

  • One-click ZIP package with the JPEG drawings, the catalogue as Excel and CSV, the box coordinates for machine learning (Excel and CSV) and a short README, plus a separate Download Excel button
  • The archive prefix is sanitized to letters, digits, _ and -
  • Export files are downloaded by the browser instead of being saved to a chosen folder

Parser Studio

  • The x at the end of a row of Current Parsed Data removes a field you assigned by mistake, together with its highlight
  • Added Examples is collapsed at the start; its title opens and closes the list, and the count stays visible
  • Previous, Next and Reset ask for confirmation when you assigned fields on the line and did not add it as an example yet
  • The pill that shows the step being processed and the box with the recognized text labels are now spaced apart in the Process OCR tab

Interface and stability

  • New look shared with the other PyPottery applications, with locally served fonts
  • The app now shuts down by itself a few seconds after you close its browser tab
  • Crashes are reported with a timestamp and the name of the thread that failed, which makes problems easier to diagnose
  • The model download shows its progress live (it stayed at 0 MB with recent versions of huggingface_hub) and no longer fails with “No such file or directory” when the app is installed in a deeply nested folder on Windows

Version 0.2.0 (Initial Release, December 2025)

🚀 A PyPotteryScan major release 🚀

  • Initial release of PyPotteryScan.
Note

The number above is the label the app showed at the time. The automatic releases that came later count from v0.1.0, which is why the published versions are 0.1.x.