A community distribution of Resemble AI's DramaBox — expressive, prompt-driven TTS with voice cloning. The prompt controls speaker identity, emotion, delivery, laughs, sighs, pauses and transitions; a 10-second voice reference clones the target timbre.
Built for Chinese users. Optimized for low-VRAM GPUs. Batteries included.
| Feature | Upstream | DramaBox Studio |
|---|---|---|
| VRAM peak | ~24 GB | ~8 GB (8 GB cards supported) |
| VRAM config | Manual | Auto-detect (3 tiers) |
| UI | Default Gradio | Midnight Gold dark cinematic theme |
| Voice library | None | Upload / name / list / delete |
| Dialogue workshop | None | Script → auto-parse characters → batch generate |
| Prompt examples | 2 English | 8 Chinese-scene examples with English direction |
| Prompt helpers | None | One-click speaker tags + 18 non-verbal cues + quote wrapper |
| Language | English prompts | English direction + Chinese dialogue supported |
demo.mp4
- Windows 10 / 11
- NVIDIA GPU with 8 GB+ VRAM
- NVIDIA driver installed
Download the all-in-one package (embedded Python + models included):
Win系统一键启动整合包，解压即用。含嵌入式 Python 环境 + 全部模型权重。
Unzip, double-click DramaBoxStudio.bat, open http://localhost:7860.
Requirements
- Python 3.10 – 3.13
- Git
- ~25 GB free disk space (for model weights)
- Internet connection (first run only — downloads model weights)
Step 1: Clone
git clone https://github.com/wwzhifeng/DramaBoxStudio.git
cd DramaBoxStudioStep 2: Install PyTorch with CUDA
pip install torch==2.8.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128Pick the CUDA version matching your driver. Check with nvidia-smi (top-right corner).
| NVIDIA driver | CUDA version | pip index |
|---|---|---|
| >= 572.00 | CUDA 12.8 | cu128 |
| >= 550.00 | CUDA 12.4 | cu124 |
| >= 545.00 | CUDA 12.1 | cu121 |
Step 3: Install remaining dependencies
pip install -r requirements.txtStep 4: Launch
python app.pyOpen http://localhost:7860 in your browser.
Option 1 users (Integration Package): models are included — skip this section.
Option 2 users (Run from Source): download the following models and place them under models/:
| Model | Source | Where to put it |
|---|---|---|
DiT Transformer (dramabox-dit-v1.safetensors) |
ResembleAI/Dramabox | models/checkpoints/ |
Audio Components (dramabox-audio-components.safetensors) |
ResembleAI/Dramabox | models/checkpoints/ |
| Text Encoder (Gemma 3 12B, 4-bit) | unsloth/gemma-3-12b-it-bnb-4bit | models/gemma/ |
Total: ~16 GB.
If HuggingFace is slow in your region, use a mirror:
set HF_ENDPOINT=https://hf-mirror.comThen download the files manually or with huggingface-cli download.
<speaker description in English>, "<dialogue>" <action direction> "<more dialogue>"
Inside quotes — spoken aloud:
- Dialogue:
"Hello, how are you?" - Vocal sounds (one word):
"Hahaha""Hehehe""Mmmmm""Ugh"
Outside quotes — performed, not spoken:
She sighs deeply.·A long pause.·He clears his throat.Her voice cracks.·He gulps nervously.·She drops to a whisper.
Never put inside quotes: Sigh Gasp Cough Ahem Pfft — the model will speak them literally.
Tips:
- Match speaker description gender/age to the voice reference
- Break long dialogue with stage directions between segments
- Use the built-in cue dropdown (18 non-verbal actions) to insert at cursor
- Select Chinese text and click
「」→""to wrap it in quotes
| Parameter | Default | Effect |
|---|---|---|
cfg_scale |
2.5 | Lower = more natural, higher = more text-faithful |
stg_scale |
1.5 | Skip-token guidance for expressiveness |
duration_multiplier |
1.1 | 10% breathing room on auto-estimated length |
steps |
30 | Euler flow matching |
| Speed | 0.5–2.0x | Tempo change without pitch shift |
- Structured prompt builder (visual fill-in-the-blank UI)
- Streaming audio playback
- Batch export with subtitle (.srt) generation
- More Chinese reference voices
- Chinese LoRA fine-tune for native-level Mandarin prosody
Contributions are welcome! This is a community project — if you have ideas, bug fixes, or want to add features:
- Bug report / Feature request: Open an Issue
- Code contribution: Fork the repo, create a branch, open a Pull Request
- Share your prompts: Found a great prompt recipe? Share it in Discussions
- Chinese voices: Have a clean 10s+ Chinese voice sample? Contributions to the voice library are especially welcome
If you're not sure where to start, check the Roadmap above or look for issues tagged good first issue.
This project is a derivative of ResembleAI/DramaBox, which is an IC-LoRA fine-tune of Lightricks/LTX-2.3. Distributed under the LTX-2 Community License Agreement — see LICENSE.
Attribution:
- Original DramaBox by Resemble AI
- Base model LTX-2.3 by Lightricks
- Community modifications by Wang Zhifeng (wangzhifeng.vip)
Model weights are auto-downloaded from ResembleAI/Dramabox on HuggingFace and are not included in this repository.
This is an unofficial community distribution. For the official DramaBox, visit ResembleAI/DramaBox.
DramaBox outputs are automatically watermarked with Resemble Perth when the watermark feature is enabled. Please comply with the LTX-2 Acceptable Use Policy when using generated audio.