ComfyUI-kaola-IndexTTS2

ComfyUI custom nodes for IndexTTS-2, a state-of-the-art zero-shot text-to-speech system with advanced emotion control capabilities.

Features

✨ Zero-Shot Voice Cloning - Clone any voice with just a few seconds of reference audio

🎭 Advanced Emotion Control - Control emotions through multiple modalities:

• Audio reference (use emotional speech samples)
• 8-dimensional emotion vectors (precise control)
• Natural language descriptions (user-friendly)

🎯 Speaker-Emotion Disentanglement - Independent control over timbre and emotion

⚡ RTX 5090 Optimized - Support for Blackwell architecture with Blackwell-specific Flash-Attention compilation and cu128 alignment

Installation

Method 1: ComfyUI Manager (Recommended)

1. Install ComfyUI Manager
2. Search for "IndexTTS-2" in the Custom Nodes Manager
3. Click Install
4. Restart ComfyUI

Method 2: Manual Installation

Clone this repository into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/YOUR_USERNAME/ComfyUI-kaola-IndexTTS2.git
cd ComfyUI-kaola-IndexTTS2

Example Workflows

• example_workflow.json — Basic voice cloning and emotion control
• 05_script_dubbing.json — Multi-character script dubbing with SRT

Step 1: Install Dependencies

Install the required Python packages:

pip install -r requirements.txt

[!IMPORTANT] External Core Library: The core IndexTTS-2 library is required. You must install it separately and ensure it's in your Python path.

Step 2: Download Model Weights

Download the IndexTTS-2 model to your ComfyUI models directory:

# Using huggingface-cli
uv tool install "huggingface-hub[cli,hf_xet]"
hf download IndexTeam/IndexTTS-2 --local-dir=ComfyUI/models/IndexTTS-2

# Or using modelscope
uv tool install "modelscope"
modelscope download --model IndexTeam/IndexTTS-2 --local_dir ComfyUI/models/IndexTTS-2

Available Nodes

1. IndexTTS2 Model Loader

Loads the IndexTTS-2 model with configurable optimization settings.

Inputs:

• model_dir - Path to model directory (default: IndexTTS-2)
• use_fp16 - Enable FP16 for lower VRAM usage
• use_cuda_kernel - Enable compiled CUDA kernels
• use_deepspeed - Enable DeepSpeed acceleration

Outputs:

• model - Loaded IndexTTS2 model instance

2. IndexTTS2 Voice Clone

Basic voice cloning using standard ComfyUI AUDIO inputs.

Inputs:

• model - IndexTTS2 model from loader
• text - Text to synthesize
• spk_audio_prompt - Reference audio (Connect to LoadAudio output)
• temperature - Randomness control (0.1-2.0, default 1.0)
• top_k / top_p - Sampling filters for better naturalness
• use_random - Enable randomness in generation

Outputs:

• audio - Generated audio (Standard ComfyUI format)

3. IndexTTS2 Emotion (Audio)

Voice synthesis with separate emotion reference audio.

Inputs:

• model - IndexTTS2 model
• text - Text to synthesize
• spk_audio_prompt - Speaker reference audio
• emo_audio_prompt - Emotion reference audio
• emo_alpha - Emotion strength (0.0-1.0)
• use_random - Enable randomness

Outputs:

• audio - Generated audio

4. IndexTTS2 Emotion (Vector)

Control emotions via 8-dimensional vector.

Inputs:

• model - IndexTTS2 model
• text - Text to synthesize
• spk_audio_prompt - Speaker reference audio
• happy, angry, sad, afraid, disgusted, melancholic, surprised, calm - Emotion intensities (0.0-1.0)
• use_random - Enable randomness

Outputs:

• audio - Generated audio

Emotion Vector Format: [happy, angry, sad, afraid, disgusted, melancholic, surprised, calm]

5. IndexTTS2 Emotion (Text)

Control emotions via natural language description.

Inputs:

• model - IndexTTS2 model
• text - Text to synthesize
• spk_audio_prompt - Speaker reference audio
• use_emo_text - Auto-extract emotion from main text
• emo_text - Optional separate emotion description
• emo_alpha - Emotion strength (0.0-1.0, recommended: 0.6)
• use_random - Enable randomness

Outputs:

• audio - Generated audio

6. IndexTTS2 Script Dubbing (SRT)

Multi-character script dubbing driven by SRT subtitles. Parses an SRT script with character names, matches each line to a voice reference, and assembles the synthesized audio onto the SRT timeline.

Required Inputs:

• model - IndexTTS2 model
• script_srt - SRT format script (multiline, see format below)
• emo_alpha - Emotion strength (0.0-2.0, default 1.0)
• temperature - Randomness control (0.1-2.0, default 1.0)
• top_k - Top-K sampling (0-100, default 0)
• top_p - Top-P sampling (0.0-1.0, default 1.0)
• use_random - Enable randomness (default False)
• save_segments - Save individual emotion clips and synthesized clips as downloadable files (default False)
• segments_prefix - Filename prefix for saved segments (default "dubbing")

Optional Inputs:

• emo_audio_prompt - Emotion reference audio (auto-sliced by SRT timestamps)
• voice_1 ~ voice_7 - Up to 7 character voice reference audios
• voice_1_name ~ voice_7_name - Character names corresponding to each voice (e.g., "唐僧")

Outputs:

• audio - Full assembled dubbed audio

SRT Script Format:

Supports standard multi-line SRT format:

1
00:00:01,000 --> 00:00:03,000
唐僧：悟空，你又调皮了。

2
00:00:04,000 --> 00:00:06,500
孙悟空：师父，俺老孙冤枉啊！

Also supports compact single-line SRT:

1 00:00:01,000 --> 00:00:03,000 唐僧：悟空，你又调皮了。
2 00:00:04,000 --> 00:00:06,500 孙悟空：师父，俺老孙冤枉啊！

Character name uses Chinese colon ： or English colon : as separator.

Emotion text in parentheses — add emotion descriptions after the character name:

1
00:00:01,000 --> 00:00:03,000
唐僧(高兴的说)：悟空，快来看。

2
00:00:04,000 --> 00:00:06,000
孙悟空（愤怒）：俺老孙不服！

Both half-width () and full-width （） parentheses are supported.

Emotion Priority:

Segment Saving:

When save_segments is enabled, individual files are saved to {output}/{segments_prefix}_segments/:

• {index}_emo_{character}_{time}.wav — Emotion audio slice for each line
• {index}_tts_{character}_{time}.wav — Synthesized audio for each line

These files appear in the ComfyUI output panel for download.

[!NOTE] Core Node Compatibility: We have removed the custom Save/Load nodes to ensure 100% compatibility with ComfyUI core. Use standard LoadAudio for inputs and SaveAudio or PreviewAudio for outputs.

Usage Examples

Basic Voice Cloning

1. Add IndexTTS2 Model Loader node
2. Add IndexTTS2 Voice Clone node
3. Connect model output to voice clone input
4. Set speaker audio path and text
5. Add IndexTTS2 Save Audio to save output

Emotion Control with Audio Reference

1. Load model with IndexTTS2 Model Loader
2. Add IndexTTS2 Emotion (Audio) node
3. Provide speaker audio (for timbre) and emotion audio (for emotion)
4. Adjust emo_alpha to control emotion strength
5. Save with IndexTTS2 Save Audio

Precise Emotion Control with Vector

1. Load model
2. Add IndexTTS2 Emotion (Vector) node
3. Set individual emotion values (e.g., surprised: 0.45, others: 0)
4. Generate and save audio

User-Friendly Emotion Control with Text

1. Load model
2. Add IndexTTS2 Emotion (Text) node
3. Enable use_emo_text to extract emotion from text
4. Or provide separate emo_text for explicit emotion description
5. Set emo_alpha around 0.6 for natural results

Multi-Character Script Dubbing

1. Load model with IndexTTS2 Model Loader
2. Add IndexTTS2 Script Dubbing (SRT) node
3. Connect model output to the dubbing node
4. Add LoadAudio nodes for each character's voice reference, connect to voice_1, voice_2, etc.
5. Set voice_1_name, voice_2_name etc. to match character names in the SRT script (e.g., "唐僧", "孙悟空")
6. (Optional) Add a LoadAudio for emotion reference audio, connect to emo_audio_prompt
7. Write your SRT script in script_srt, using 角色名：台词 format
8. (Optional) Add emotion descriptions in parentheses: 唐僧(高兴的说)：悟空，快来看。
9. Enable save_segments to export individual audio clips for review
10. Connect output to SaveAudio or PreviewAudio

[!TIP] An example workflow is available at examples/05_script_dubbing.json. Import it directly into ComfyUI to get started quickly.

Tips

💡 FP16 Mode: Highly recommended for faster inference and lower VRAM usage with minimal quality loss

💡 Emotion Alpha: When using text-based emotion control, use lower emo_alpha values (0.6 or less) for more natural speech

💡 Random Sampling: Enabling use_random adds variety but may reduce voice cloning fidelity

💡 Audio Format: Reference audio should be clear, with minimal background noise

💡 Voice Cloning Accuracy: For best results, use 5-15 seconds of clean, single-speaker reference audio. Avoid background music or noise. Consistent speaking style in the reference yields more stable cloning.

💡 Script Dubbing Timeline: If a synthesized clip is longer than the gap before the next SRT timestamp, the next clip is automatically pushed forward (no truncation). The output may be longer than the original SRT timeline.

Troubleshooting

Model Not Found Error

Make sure the model is downloaded to the correct location:

ComfyUI/models/IndexTTS-2/config.yaml
ComfyUI/models/IndexTTS-2/[model files]

Protobuf builder Error

If you see ImportError: cannot import name 'builder' from 'google.protobuf.internal', your environment has a Protobuf version mismatch. Fix it with:

pip install protobuf==3.20.3

RTX 5090 / CUDA 13.0 Considerations

For the best performance on RTX 5090, ensure your Torch environment is aligned with CUDA 12.8 or higher:

pip install --force-reinstall torch==2.8.0 torchaudio==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128

Flash-Attention Compilation

If flash-attn fails on your 5090, perform a local source compilation:

pip install ninja
pip install flash-attn --no-build-isolation --no-cache-dir

Slow HuggingFace Downloads

Set mirror endpoint (for users in China):

export HF_ENDPOINT="https://hf-mirror.com"

Resources

• 📄 IndexTTS-2 Paper
• 🎬 Demo Page
• 💻 GitHub Repository
• 🤗 HuggingFace Model
• 🌐 ModelScope Model

Citation

If you use IndexTTS-2 in your work, please cite:

@article{zhou2025indextts2,
  title={IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech},
  author={Siyi Zhou, Yiquan Zhou, Yi He, Xun Zhou, Jinchao Wang, Wei Deng, Jingchen Shu},
  journal={arXiv preprint arXiv:2506.21619},
  year={2025}
}

Using Local Models (Advanced)

The official IndexTTS-2 requires internet access to download the wav2vec2bert model from HuggingFace. If you want to use a fully offline setup with local models, you can use the fork version:

Install Fork Version

# Clone the fork instead of the official repo
git clone https://github.com/kana112233/index-tts.git
cd index-tts

# Install dependencies
GIT_LFS_SKIP_SMUDGE=1 git checkout -f
pip install -e .

Download wav2vec2bert Model

# Download to a local directory
huggingface-cli download facebook/w2v-bert-2.0 \
  --local-dir /path/to/models/w2v-bert-2.0

Set Environment Variable

Before starting ComfyUI, set the local model path:

export W2V_BERT_PATH="/path/to/models/w2v-bert-2.0"

What's Different in the Fork?

The fork modifies indextts/utils/maskgct_utils.py to:

• Check for the W2V_BERT_PATH environment variable
• Use local files if the path exists
• Fall back to HuggingFace if not set

This allows completely offline usage without internet access.

License

This project follows the license of the original IndexTTS-2 project. For commercial usage, please contact: indexspeech@bilibili.com

Acknowledgements

• IndexTTS-2 - The amazing TTS model
• ComfyUI - The powerful node-based UI