DramaBox in ComfyUI: local TTS setup and voice cloning

This tutorial walks through DramaBox, Resemble AI's text-to-speech model, using community-built ComfyUI nodes. The speaker describes it as an audio-focused derivative of LTX 2.3 that avoids loading the full video model. A Python pipeline with a Gradio interface is another local execution option, though the walkthrough focuses on ComfyUI.

Setup involves cloning the custom node repository into ComfyUI's custom nodes folder and installing its requirements.txt dependencies. The speaker recommends 24 GB of VRAM: the roughly 6.5 GB DramaBox download includes the main model and audio components, while Unsloth's quantized Gemma 3 12B text encoder needs additional headroom. The model setup also requires every split Gemma weight file and its tokenizer.

After restarting ComfyUI, the demonstration connects reference audio to the DramaBox TTS node's voice sample input and previews the generated speech. Examples use a generated female voice as the cloning reference, with different dialogue and emotion prompts. The speaker reports fast generation and prefers the expression to other open source TTS options, but provides no timed benchmark or controlled comparison.

The final workflow combines DramaBox dialogue with LTX prompt relay for a short scene. Character audio durations determine frame timing. A vision language model supplies camera prompts, though the speaker also describes writing prompts manually with ComfyUI's timeline node.