Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

v0.2.1 - 2026-11-19

Added

  • An Acknowledgements section in the README: the open tools the pipeline uses, and the CC BY-NC 4.0 licence of the MMS-LID weights that the lid extra downloads.

Changed

  • The contact address is contact@burkimbia.com: in the PyPI metadata, the README and __email__.

v0.2.0 - 2026-11-19

First release on PyPI.

Added

  • VLMOCR runs on Apple's GPU (MPS) when there is no CUDA, and loads Qwen3-VL in bfloat16 there: on a Mac M5 it reads as float32 does, 30 times faster. It fell back to the CPU before.
  • strip_overlays and process_batch(overlay_frames=60) take logos and banners out of the frames before merging; mark_non_subtitles marks credits, the video's title and scrolling text in the merged segments.
  • windows joins subtitles less than 2 s apart into windows of 5 to 30 s, as segments AudioGenerator cuts.
  • LanguageID (MMS-LID, the new lid extra) says which language each window is spoken in; written_in says whether a subtitle is written in French or English.
  • TextIndex (the new copies extra) finds which video already read a subtitle line comes from, to tell a copy published by another channel.
  • A pipeline diagram drawn in draw.io (docs/diagrams/pipeline.drawio) in the README and the guide, and a README example that runs the whole pipeline.
  • scripts/figures.py makes the guide's figures from sandbox/sample_short.mp4 with the library itself.

Security

  • certifi 2026.10.23 in uv.lock, and uv pinned to 0.13.4 in CI and in the pre-commit hook: earlier versions trust the compromised emSign ECC Root CA - G3 root.

Changed

[0.1.0] - 2026-10-06

Not published on PyPI.

Added

  • pyproject.toml, built with hatchling, and uv.lock for development. The OCR engines are extras: easyocr, paddle, vlm, openai, all.
  • VLMOCR reads with any image-text-to-text model of transformers, Qwen3-VL-2B by default: on 48 subtitle frames it made 0.53 % character error, against 37.49 % for tool="easyocr".
  • merge_segments: readings of one subtitle are merged when they share most of their words, one empty frame does not end a subtitle, and each segment has best_time, the frame read best.
  • OCRManager(region=...): the subtitle band, detected once per video when it is not given.
  • YoutubeConsumer(client=...): the pytubefix client, MWEB by default; files are named youtube_<id>.mp4.
  • A documentation site: https://sawallesalfo.github.io/Frame2Text4LLM/.

Fixed

  • A segment ends when its subtitle leaves the screen; a subtitle seen on one frame no longer lasts 0 s.
  • AudioGenerator reads the frames' MM:SS:mmm times and seeks before reading: audio was cut in the wrong place.
  • SubtitleRegionDetector takes the full width of the image; its margin cut the first letters.
  • OCRManager loads a model once, not once per image.
  • SRTExporter writes valid times past an hour, and each subtitle's own end.

[0.0.1]

  • First version, not published.