guey-khala-mari/LTX-2.3-Resemble-Ai-Dramabox-IC-LoRa

DramaBox LoRA (LTX-2.3 audio) — extracted

A standalone LoRA extracted from resemble-ai/DramaBox, so it can be applied to a base LTX-2.3 3.3B audio-only model in a normal LoRA workflow instead of running DramaBox's merged 6.6 GB checkpoint + TTS wrapper.

DramaBox is Resemble AI's expressive TTS: an IC-LoRA fine-tune of the LTX-2.3 audio branch. The prompt controls speaker identity, emotion, delivery, laughs, sighs and pauses; an optional voice-reference clip clones timbre.

What this is

• The DramaBox fine-tune delta only — the audio-branch attention / feed-forward layers (audio_attn1, audio_attn2, audio_ff).
• Extracted against the LTX-2.3 dev audio checkpoint. Use the dev base, not the distilled one.
• Everything else (Gemma-3-12B text encoder, audio VAE, BigVGAN BWE vocoder) is unmodified stock LTX-2.3 — not included here.

Usage (ComfyUI)

1. Load the LTX-2.3 dev audio model (UNETLoader).
2. Apply this LoRA with LoraLoaderModelOnly.
3. Text-encode with the Gemma-3-12B LTX audio text encoder. Negative prompt (needed at CFG > 1): worst quality, inconsistent, robotic, distorted, noise, static, muffled, unclear, unnatural, monotone
4. Sample: 30 steps, Euler, LTX2Scheduler sigmas. Guidance: cfg 7.0, STG 1.0 @ block 29, rescale ≈ 0.8.
5. Decode with the LTX-2.3 audio VAE (confirm output ≈ 48 kHz — BWE active).

For faithful voice-reference conditioning and prompt-driven clip length, the bundled nodes/ComfyUI-LTXV-ResembleAI-Dramabox node pack adds:

• DramaBox Voice Preprocess — mono, trim/pad, −4 dBFS
• DramaBox Apply Voice Reference — overlapping RoPE positions + asymmetric read-only attention mask (matches DramaBox training; native LTX ID-LoRA conditioning uses a different convention)
• DramaBox Estimate Audio Duration — prompt → frame count

Prompting

Quoted text is spoken literally; unquoted text is stage direction. Use generic speaker nouns (A weary man, A nervous young woman) — never roles.

A man speaks evenly, "I gave you one job."
His voice rises with fury, "AND YOU MESSED IT UP!"

See the DramaBox prompting guide.

Not included / not applied

• Perth watermarking — DramaBox watermarks its output by default. This LoRA does not. Add resemble-perth downstream if you need it.
• RE-USE reference denoise (NVIDIA nvidia/RE-USE, NSCLv1 non-commercial).
• Long-form chunk-and-stitch, end-of-clip silence fix.

License

Distributed under the LTX-2 Community License Agreement — see LICENSE. As a derivative of LTX-2, this LoRA cannot be relicensed. The agreement's attribution, use-restriction and revenue-threshold terms apply and pass through to anyone you distribute it to.

Credits

• Resemble AI — DramaBox
• Lightricks — LTX-2 base model