ComfyUI-IndexTTS2

Lightweight ComfyUI wrapper for IndexTTS 2 (voice cloning + emotion control). The nodes call the original IndexTTS2 inference and keep behavior faithful to the repo.

Original repo: https://github.com/index-tts/index-tts

Install

• Clone this repository to: ComfyUI/custom_nodes/
• In your ComfyUI Python environment:

  pip install wetext
  pip install -r requirements.txt
  

Models (checkpoints)

• Create a folder named 'checkpoints' in the root directory

• Download ALL files and subfolders from Hugging Face and put them under the new 'checkpoints' folder, preserving the original structure: https://huggingface.co/IndexTeam/IndexTTS-2/tree/main

  Optional, if auto-cached online if missing:

• Additional required files for local loading (download these separately):

  • W2V-BERT-2.0 feature extractor/model (download from: https://huggingface.co/facebook/w2v-bert-2.0):
    • Download the entire repository contents and place them under: checkpoints/w2v-bert-2.0/
  • BigVGAN files (download from: https://huggingface.co/nvidia/bigvgan_v2_22khz_80band_256x):
    • Download file: config.json → place in: checkpoints/bigvgan/
    • Download file: bigvgan_generator.pt → place in: checkpoints/bigvgan/
  • Semantic codec (download from: https://huggingface.co/amphion/MaskGCT/tree/main):
    • Download file: semantic_codec/model.safetensors → place in: checkpoints/semantic_codec/
  • CAMPPlus model (download from: https://huggingface.co/funasr/campplus/tree/main):
    • Download file: campplus_cn_common.bin → place in: checkpoints/

• Complete checkpoints folder structure:

  ComfyUI/custom_nodes/ComfyUI-IndexTTS2/checkpoints/
  ├── config.yaml
  ├── gpt.pth
  ├── s2mel.pth
  ├── bpe.model
  ├── feat1.pt
  ├── feat2.pt
  ├── wav2vec2bert_stats.pt
  ├── campplus_cn_common.bin
  ├── bigvgan/
  │   ├── config.json
  │   └── bigvgan_generator.pt
  ├── semantic_codec/
  │   └── model.safetensors
  ├── qwen0.6bemo4-merge/          (required only for Text -> Emotion node)
  │  └── [all Qwen model files]
  └── w2v-bert-2.0/
      └── [all bert files]
  

Important: The updated code now uses local model files by default for offline usage and faster loading.

Nodes

• IndexTTS2 Simple

  • Inputs: audio (speaker), text, emotion_control_weight (0.0-1.0), emotion_audio (optional), emotion_vector (optional)

  • Outputs: AUDIO (for Preview/Save), STRING (emotion source message)

  • Notes: device auto-detected, FP16 on CUDA, 200 ms pause between segments (fixed), emotion precedence = vector > second audio > original audio

• IndexTTS2 Emotion Vector

  • 8 sliders (0.0-1.4) for: happy, angry, sad, afraid, disgusted, melancholic, surprised, calm
  • Constraint: sum of sliders must be <= 1.5 (no auto-scaling)
  • Output: EMOTION_VECTOR

• IndexTTS2 Emotion From Text (optional)

  • Input: short descriptive text
  • Requires: modelscope and local QwenEmotion at checkpoints/qwen0.6bemo4-merge/
  • Outputs: EMOTION_VECTOR, STRING summary

Examples

• Basic: Load Audio -> IndexTTS2 Simple -> Preview/Save Audio
• Second audio emotion: Load Audio (speaker) + Load Audio (emotion) -> IndexTTS2 Simple -> Save
• Vector emotion: IndexTTS2 Emotion Vector -> IndexTTS2 Simple -> Save
• Text emotion: IndexTTS2 Emotion From Text -> IndexTTS2 Simple -> Save

Troubleshooting

• Tested only in Windows. DeepSpeed disabled.
• Emotion vector sum exceeds maximum 1.5: lower one or more sliders or adjust the text-derived vector.
• BigVGAN kernel message: custom CUDA kernel is disabled by default; falls back to PyTorch ops.
• Missing 'wetext' module: Run pip install wetext to fix this Windows-specific dependency.
• 404 Repository Not Found errors: Ensure all additional model files are downloaded to your checkpoints folder as described above.
• Model loading issues: Verify your checkpoints folder contains all required files with the correct directory structure.

Expected Output: When working correctly, you should see messages like:

• Loading config.json from local directory
• Loading weights from local directory
• All model paths pointing to your local checkpoints folder

Performance: The system processes audio through 4 stages (Text → GPT → S2Mel → BigVGAN). Multiple progress bars and tensor size outputs are normal during inference.