@auriodk

Version 3.0.0. I'm having repeated problems with music-video planning in Maestro. The LLM/Director fails to produce valid shot plans, even after multiple repair attempts. The errors include missing or incorrectly ordered source events (M1, M2, etc.), misinterpreting simple instructions such as "both hands," and failing to assign every source event exactly once in the correct order. These failures happen during planning, before video generation even begins. I've simplified and rewritten my prompts several times, but the problems persist. It seems the planning and validation system is struggling with its own requirements.

Fixes to h3_dialogue.py. I had too many duplicates and also some wrong hair descriptions, the latter H3 simply can´t handle, no point in describing "short dark hair" as an attribute, "short dark hair" probably has thousands of variations and H3 invents its own. This is a brief summary copied from ChatGPT of what it did to the file: 1. Fixed duplicate-subject mapping so an existing character like Don/Maria is matched to <Picture 1>/<Picture 2> instead of creating new Subject 3/4 copies. 2. Fixed prompt recompilation buildup, stripping compiler-generated identity and lip-sync boilerplate before recompiling so it doesn’t keep duplicating itself. 3. Added hair to picture identity, so the reference now carries “facial, hair, bodily, and character identity” from the picture. 4. Updated the cleanup regex so it recognizes both the old and new identity wording. Syntax-checked the finished file afterward. I ran a difficult test of 50 shots afterwards and got 1 duplicate due to poor continuity description in my prompt in the last shot, it seems fixed for the most part.