⚙️ DramaBox Engine

If you've ever wished your TTS would act a line instead of reading it like a GPS, DramaBox is the engine you've been waiting for - and it's also the one that'll make you check your VRAM counter twice. It's Resemble AI's 3.3B expressive TTS model, and it's the closest thing in open weights to an actual director's chair. You write prose around quoted dialogue - stage directions, delivery, even a sigh - and the model performs it.

DramaBox is one of the engines inside TTS Audio Suite (diodiogod's catch-all TTS pack, evolved from the old ChatterBox SRT project). This node configures the engine; its single TTS_engine output feeds the suite's Unified TTS Text or Unified SRT TTS nodes, where the actual generation happens. You don't call anything here - this is pure setup.

How it works

Under the hood DramaBox is an LTX-family diffusion model for audio, and it brings diffusion-adjacent controls with it. It uses CFG (text guidance) plus STG (skip-token guidance) to follow your prompt, a negative prompt to discourage artifacts, and a duration estimator that predicts how long each line should take. It's English-only, and the reference window matters: connect a voice reference and the first ref_duration seconds (default 10) are what get cloned.

The prompt_template field is the secret sauce. Default is "{seg}" - every plain line is treated as spoken dialogue. Change it to something like A man speaks warmly, "{seg}" and every segment picks up that delivery. That's how you get a whole audiobook narrated in one consistent mood without editing per line.

The inputs that actually matter

You'll touch maybe four of these:

• cfg_scale (2.5) - higher is more text-faithful, lower is more natural. If output sounds forced, drift it down toward 2.
• stg_scale (1.5) - skip-token guidance. Leave it alone until you know what you're doing; it's not the knob you think it is.
• memory_mode - fast keeps everything on CUDA and targets roughly 24GB VRAM. staged and sequential are experimental ways to fit less VRAM by reloading components per segment. Real, but slow.
• ref_duration - seconds of your reference audio that get used. A clean 10-second single-speaker clip beats a 30-second one with noise.

gen_duration should stay at 0 (automatic estimation) for normal work - it's for explicit per-segment durations. transformer_quantization offers fp8_cast for lower VRAM at some speed cost, compile_model trades a very slow first generation for faster later ones, and the local_lora_adapter / lora_strength inputs load a trained DramaBox LoRA right here. The seed isn't on this node - the unified TTS nodes supply it.

Installing it

Install the pack once, then this node just works:

cd ComfyUI/custom_nodes
git clone https://github.com/diodiogod/TTS-Audio-Suite.git
cd TTS-Audio-Suite
python install.py   # or ../../../python_embeded/python.exe install.py on portable

ComfyUI Manager works too - search "TTS Audio Suite". The installer sorts out Python 3.12+/3.13 compatibility and dependency conflicts. On first run the model auto-downloads roughly 16.4GB into ComfyUI/models/TTS/dramabox/DramaBox/ (including the bundled 4-bit Gemma text encoder), so budget disk and patience. You need a CUDA GPU; CPU inference is unsupported.

Where people get burned

VRAM. The community consensus is that 16GB runs it only with memory tricks - people reach for staged mode, fp8_cast, and offloading to keep it alive. If it OOMs, that's the first thing to try, not a bug report.

Near-silent output. DramaBox occasionally returns a near-silent clip for a particular reference/duration/seed combination. The suite detects it and prints a warning - preserve the audio, try a different seed or reference window, and move on.

The license. DramaBox ships under the LTX-2 Community License. Free under USD $10M annual revenue (all revenue, not profit, and affiliates count), paid above that. Fine for hobbyists and small studios; not free forever.

The first generation also downloads and loads Gemma, so it feels like the node is broken. It isn't. Give it a minute.