ComfyUI custom nodes for DramaBox — ResembleAI's expressive text-to-speech system built on the LTX-2.3 audio diffusion transformer.

Please Note: DramaBox offers two generation methods. dramabox_wrapper runs the original DramaBox pipeline for native behavior, while clip_loader uses ComfyUI memory management for a more integrated and reliable workflow. Users wanting to experiment, can enable wrapper mode in Preferences or in the DramaBox Options node.

VRAM Recommendation: For the smoothest experience, a 16 GB GPU is strongly recommended. Lower-VRAM GPUs can still work with aggressive offloading, but may be slower and more prone to out-of-memory errors.

Users with sufficient VRAM can disable model offloading in Preferences for faster generation.

DramaBox uses a Gemma 3 12B text encoder. By default the node loads gemma_3_12B_it_fp4_mixed.safetensors — the same file used by ComfyUI's own LTX-2 workflows, so if you already have it you're good to go. If it is not present it is downloaded automatically into your ComfyUI/models/text_encoders/ folder on first use.

Per-installation preference — open ComfyUI Settings → DramaBox → Default Text Encoder filename and enter a Gemma filename from your models folder. Supports .safetensors and .gguf (with ComfyUI-GGUF installed). If you omit the extension, DramaBox tries .safetensors first, then .gguf. Leave it blank to keep the fp4 default.

Memory preference — open ComfyUI Settings → DramaBox → Memory and keep Automatic model offload (text encoder + post-generate) enabled (default). This does two things automatically:

• Offloads Gemma right after prompt encoding.
• Offloads DramaBox models to CPU after generation.

If you have plenty of VRAM and prefer maximum throughput, you can disable this preference to keep models resident on GPU.

Per-workflow override — add a DramaBox CLIP Loader node, select the model you want (.safetensors or optional .gguf), and connect its output to the TTS node's dramabox_clip input. This takes precedence over the global preference and lets you switch encoders between workflows without touching settings.

The DramaBox Options node includes post_generate_model_policy with three modes:

• keep_loaded — keep Gemma and DramaBox on GPU (fastest next run, highest VRAM usage).
• offload_to_cpu — offload Gemma after text encoding and offload DramaBox to CPU after generation. (Default)
• offload — offload Gemma after text encoding and fully unload DramaBox after generation.

For most users, offload_to_cpu is a good balance between memory savings and iteration speed.

Set via DramaBox Options (generation_mode) or globally in ComfyUI Settings → DramaBox → Use DramaBox Wrapper mode by default.

Advantages:

• Best ComfyUI-native VRAM control.
• Text encoding and diffusion are staged cleanly (less overlap in VRAM).
• Best fit when using custom CLIP Loader / GGUF text encoders.

Disadvantages:

• Not the original native DramaBox warm-server path.
• Output can differ from native DramaBox (tone, pacing, or phrasing may vary).

Advantages:

• Closest to original DramaBox warm-server behavior.
• Best match when you want native DramaBox-like output.

Disadvantages:

• In keep_loaded, prompt encoder and transformer can overlap in VRAM.
• ComfyUI cannot fully manage wrapper internals like it does patcher-based models.

Wrapper + memory policy notes:

• keep_loaded: fastest repeated runs, highest VRAM use.
• offload_to_cpu / offload: uses one-shot low-memory execution (including LoRA entries) to avoid warm overlap, then releases wrapper cache.
• Add DramaBox Unload to force cleanup at a specific workflow point.

The DramaBox TTS node accepts a lora_stack input (connect any ComfyUI LORA_STACK output). LoRA weights are applied directly into the already-loaded model and removed immediately after generation, so you can switch LoRAs between runs without triggering a slow model reload.

LoRAs and voice reference samples work independently and can be used together. A LoRA bakes a trained voice style into the model weights, while a voice reference sample is fed as audio conditioning during generation. Using both at once — for example a LoRA trained on a voice alongside a reference clip of that same voice — will reinforce the effect.

Note: DramaBox LoRAs are specific to this model and cannot be used with other ComfyUI nodes such as LTX Video.

Voice LoRAs for DramaBox can be trained with Voice Clone Studio — DramaBox Edition, a stripped-down version of Voice Clone Studio, made for DramaBox, for both inference and LoRA training. It provides a dataset creator that automatically transcribes and splits your long audio clips into smaller clips ready to be used by the trainer. Drop the resulting .safetensors file into ComfyUI's models/loras/ folder.

Follow these steps:

Open a terminal in your ComfyUI custom_nodes folder, then run:

git clone https://github.com/FranckyB/ComfyUI-DramaBox.git

Choose one option below that matches your ComfyUI install.

1. Open your ComfyUI_windows_portable folder.
2. Open a terminal in that folder.
3. Run:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-DramaBox\requirements.txt

(If your folder is named python_embedded, use that instead of python_embeded.)

1. Open the folder where ComfyUI Desktop is installed.
2. Open a terminal in that folder.
3. Run (adjust the path if your layout differs):

python\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-DramaBox\requirements.txt

1. Open a terminal in your ComfyUI folder.
2. Activate the same environment you use to launch ComfyUI.
3. Install the requirements:

Windows (cmd):

cd C:\path\to\ComfyUI
venv\Scripts\activate
pip install -r custom_nodes\ComfyUI-DramaBox\requirements.txt

Linux / macOS (bash/zsh):

cd /path/to/ComfyUI
source venv/bin/activate
pip install -r custom_nodes/ComfyUI-DramaBox/requirements.txt

After restart, search for DramaBox TTS in the node menu.

• If install fails, check that the requirements path is correct.
• If ComfyUI starts but node is missing, confirm folder is custom_nodes/ComfyUI-DramaBox (not nested like ComfyUI-DramaBox-main/ComfyUI-DramaBox).
• If pip command installs to the wrong place, you are likely using a different Python than ComfyUI. Re-run using option A, B, or your active venv Python.

• Generation mode control — added generation_mode (clip_loader or dramabox_wrapper) in DramaBox Options plus a global default wrapper-mode preference.
• Wrapper LoRA parity — wrapper mode now supports direct LoRA stack/strength application and cleanup, matching non-wrapper behavior.
• Wrapper offload behavior — when wrapper mode uses offload_to_cpu / offload, it now runs through one-shot low-memory execution (with LoRA support) to reduce VRAM overlap.
• DramaBox Unload node — added a passthrough unload node to release both clip and wrapper caches on demand.
• Text encoder overhaul — DramaBox now uses ComfyUI's standard CLIP infrastructure for the Gemma text encoder, matching the native LTX-2 loading path for correct VRAM management.
• Default encoder — switched to gemma_3_12B_it_fp4_mixed.safetensors (ComfyUI/Comfy-Org's own quantized file, ~8 GB vs ~24 GB for the previous bnb-4bit snapshot). Downloaded automatically into text_encoders/ on first use if not already present.
• DramaBox CLIP Loader node — optional node to load Gemma text encoders from text_encoders/ (safetensors) and GGUF files when ComfyUI-GGUF is installed. Connect to the TTS node's dramabox_clip input for per-workflow encoder selection.
• Settings preference — added ComfyUI Settings → DramaBox → Default Text Encoder filename to set a global default without needing a CLIP Loader node in every workflow.
• LTX package compatibility — DramaBox now relies on installed ltx-core / ltx-pipelines packages and ships only DramaBox-specific compatibility code in py/, reducing conflicts with other LTX-based custom nodes.
• Old Gemma snapshot cleanup — the large gemma-3-12b-it-bnb-4bit/ model directory (previously downloaded into models/dramabox/) is automatically removed on startup since it is no longer needed.
• Removed info output — the info string output has been removed from the DramaBox TTS node.

• DramaBox by ResembleAI
• Built on the LTX-Video audio diffusion architecture by Lightricks