ComfyUI-IndexTTS2

Lightweight ComfyUI wrapper for IndexTTS 2 (voice cloning + emotion control). Nodes call the upstream inference code so behaviour stays matched with the original repo.

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

Updates

• 2025-09-22: Added IndexTTS2 Advanced node exposing sampling, speed, seed, and other generation controls.

Install

• Clone this repository into ComfyUI/custom_nodes/
• Inside your ComfyUI Python environment:

  pip install wetext
  pip install -r requirements.txt
  

Models

• Create checkpoints/ in the repo root and copy the IndexTTS-2 release there (https://huggingface.co/IndexTeam/IndexTTS-2/tree/main). Missing files will be cached from Hugging Face automatically.

Nodes

• IndexTTS2 Simple – speaker audio, text, optional emotion audio/vector; outputs audio + status string. Auto-selects device, FP16 on CUDA.
• IndexTTS2 Advanced – Simple inputs plus overrides for sampling, speech speed, pauses, CFG, seed.
• IndexTTS2 Emotion Vector – eight sliders (0.0–1.4, sum <= 1.5) producing an emotion vector.
• IndexTTS2 Emotion From Text – requires ModelScope and local QwenEmotion; turns short text into an emotion vector + summary.

Examples

• Speaker audio -> IndexTTS2 Simple -> Preview/Save Audio
• Speaker + emotion audio -> IndexTTS2 Simple -> Save
• Emotion Vector -> IndexTTS2 Simple -> Save
• Emotion From Text -> IndexTTS2 Simple -> Save

Troubleshooting

• Windows only so far; DeepSpeed is disabled.
• Install wetext if the module is missing on first launch.
• If w2v-bert keeps downloading, confirm checkpoints/w2v-bert-2.0/ exists (or set W2V_BERT_LOCAL_DIR).
• 404 or load failures usually mean a missing file in checkpoints/; re-check the tree above.
• Emotion vector sum must stay <= 1.5.
• BigVGAN CUDA kernel warnings are expected; PyTorch fallback kicks in automatically.

Logs you should see

• Loading config.json from local directory
• SeamlessM4TFeatureExtractor loaded from: checkpoints/w2v-bert-2.0/
• Model paths pointing at your checkpoints/ tree.