Switch Gemma to google/gemma-3-4b-it (CPU-compatible, no bitsandbytes needed)

Revert inference_server.py to original (undo bad dtype indentation)

Fix IndentationError on line 103: correct fp32 dtype assignment

Add fp32 dtype support for CPU compatibility

Add explicit Target Duration slider (0–60s) + gen_duration kwarg

Revert: keep DramaBox naming (rebrand reverted per CEO)

Rebrand: DramaBox → LTX-2.3-Voice

Patch end-of-clip silence prior at latent frame 513 for >20s outputs

Use richer duration estimator in warm server (sentence + non-verbal aware)

Fetch unsloth bnb-4bit Gemma + Dramabox checkpoints from HF on startup

DramaBox Space — initial app + vendored ltx2