#!/usr/bin/env python3
"""
Minimal inference example for LAION-Box Emotional v0.7.

These are fully-merged standalone DiT checkpoints — use them as the `--checkpoint`
(transformer/DiT) weights in the DramaBox / LTX-2 audio pipeline. You still need the
audio-components checkpoint (ltx-2.3-22b-dev.safetensors) and the Gemma prompt encoder.

    pip install torch safetensors librosa soundfile   # + the DramaBox/ltx2 pipeline
"""
import sys

# 1) point at the DramaBox pipeline source
sys.path.insert(0, "DramaBox/src")
from inference_server import TTSServer

MODEL          = "LAION-Box-Emotional-v0.7_best1_step850.safetensors"  # this repo
AUDIO_COMPONENTS = "dramabox-audio-components.safetensors"     # VAE + vocoder (in this repo)
GEMMA_ROOT     = "/path/to/gemma"                                     # prompt encoder

tts = TTSServer(
    checkpoint=MODEL,
    full_checkpoint=AUDIO_COMPONENTS,
    gemma_root=GEMMA_ROOT,
    device="cuda", dtype="bf16", bnb_4bit=True,
)

tts.generate_to_file(
    prompt="A woman, voice breaking with emotion: 'After all these years... you finally came back.'",
    output="emotional_out.wav",
    voice_ref="reference_voice.wav",   # 5-10s clean speaker sample
    cfg_scale=2.5,      # higher -> follows the emotional prompt harder
    stg_scale=1.5,
    duration_multiplier=1.1,
    ref_duration=10.0,
    denoise_ref=True,
    seed=42,
)
print("wrote emotional_out.wav")