v2.3.0

This release removes the two-pass subtitle generation system in favor of a simpler, single-pass workflow. The second "QA" pass proved to be of limited value and added unnecessary complexity and cost.

BREAKING CHANGES:

• Configuration:
  • The ai.pass1_model and ai.pass2_model settings have been removed and replaced with a single ai.model_subtitles.
  • The ai.lyrics_model setting has been renamed to ai.model_lyrics.
  • The thread.subtitles1 and thread.subtitles2 settings have been consolidated into thread.subtitles.
  • The ai.model shorthand now sets model_subtitles and model_lyrics.
• State Persistence: The format for intermediate job state files has changed.
  • part_XXX.pass1.MODEL.json is now part_XXX.subtitles.MODEL.json.
  • part_XXX.pass2.MODEL.json is no longer created.
  • You must delete temporary directories (e.g., tmp_<video_name>) from previous runs before using this version.

Improvements:

• Simplified Pipeline: The subtitle generation process is now a single AI call, reducing complexity, processing time, and potential points of failure.

• Prompt Versioning & Auto-Reprocessing:

  • Introduced versioning for both the "Lyrics/Scene Detection" prompt (LYRICS_PROMPT_VERSION) and the "Subtitle Generation" prompt (SUBTITLES_PROMPT_VERSION).
  • The application now automatically detects when a prompt has been updated between runs. It will invalidate cached results for affected segments and re-process them using the new prompt, ensuring subtitles always reflect the latest logic.
  • Cache loading is now more robust. Corrupted or outdated job files from previous runs are automatically ignored and re-processed, preventing crashes.

• Prompt Engineering:

  • Updated SUBTITLES_PROMPT_VERSION to 10.
  • The subtitle generation prompt has been significantly improved with a "Decoding Hierarchy" to better handle ambiguous audio by prioritizing on-screen text and reference JSON.
  • Added stricter rules to prevent "cascading delay" timing errors and to enforce intelligent segmentation.