ComfyUI-MelodramaBox

Why this fork exists

This is a fork, maintained by Mad Pony Interactive for use in Cubric Vision. It is based on comfyui-melodramabox 2.1.0 from the ComfyUI Registry.

The registry release names github.com/doggeddalle/ComfyUI-MelodramaBox as its source repository, but that URL 404s — there is no upstream repository to open a pull request against. Cubric Vision installs its ComfyUI nodes by pinned git commit, and it needs three bug fixes that the published zip does not have, so the only way to ship them was to publish a fork.

The three fixes: the shared text-encoder root is honoured instead of only the first one (otherwise ~8 GB of Gemma-3 is re-downloaded over a copy that is already there), the documented .gguf DiT path is actually selectable, and the unload node stops being cached out — it exists purely for its side effect, so a cached run silently did nothing. Each is marked in place with an MPI-607 comment and described in NOTICE.

All credit for the pack is doggeddalle's. The Apache-2.0 licence is unchanged. If the upstream repository reappears, these fixes are offered back gladly.

Expressive, prompt-driven text-to-speech with zero-shot voice cloning for ComfyUI — a native implementation of Resemble AI's DramaBox (the LTX-2.3 3.3B audio-only DiT by Lightricks). Audio only.

Write what you want said and how it's performed — the prompt controls speaker identity, emotion, delivery, laughs, sighs, and pauses. Add a ~10-second reference clip to clone a target voice, with no training needed.

A woman speaks warmly, "Hello — it's good to finally meet you." She laughs softly.

Why this pack

• Native ComfyUI, not a wrapper. Each model component (DiT, Gemma-3 text encoder, audio VAE + vocoder) is a real ModelPatcher, so ComfyUI manages VRAM, offloading, and eviction for you — with a live progress bar and Cancel.
• Tuned for real GPUs. Runs comfortably on a 24 GB card; 12 GB cards work via offload, and lower via GGUF.
• GGUF quantization of the audio DiT — smaller files, less VRAM (details).
• Zero-shot voice cloning from a short reference clip (guide).
• 48 kHz output, long-form generation, and Perth watermarking.
• Use other LTX-2.3 finetunes by extracting their audio branch (details).

Install

cd ComfyUI/custom_nodes
git clone <this repo> ComfyUI-MelodramaBox
pip install -r ComfyUI-MelodramaBox/requirements.txt

Restart ComfyUI. Model weights (~16 GB total) download automatically the first time you run a loader. Full details, including the model list and the bf16 alternative to bitsandbytes, are in the Guide.

The graph

DiT Loader ─────────────────┐   (audio DiT; .safetensors or .gguf)
Text Encoder Loader → Text Encode ─┤   (Gemma-3 12B; 4-bit or bf16)
Audio VAE Loader ─┬─ Voice Reference ─┤   (optional cloning clip)
                  │                    ▼
                  │             Sampler → VAE Decode → Save Audio
                  └───────────────────────┘

Start from a ready-made workflow in example_workflows/: dramabox_standard_inference.json (24 GB), dramabox_lowvram_12gb.json (12 GB), or dramabox_advanced_inference.json (long-form + LoRA + voice ref).

VRAM at a glance

Measured on an RTX 3090. See Advanced → VRAM management.

Documentation

• Guide.md — install, every node, voice cloning, quality controls, LoRA loading.
• Advanced.md — architecture, VRAM tuning, GGUF, using other LTX-2.3 finetunes, troubleshooting, Windows RE-USE setup.
• conversion_scripts/ — standalone tools to turn any LTX-2.3 finetune into a small GGUF audio DiT.

License

This pack's own code is Apache 2.0 (see LICENSE). The DramaBox model and the vendored LTX-2 inference code are under the LTX-2 Community License; optional RE-USE reference denoising is NSCLv1 (non-commercial). See Advanced → License for the full breakdown.