Nodes/ComfyUI-VibeVoice/VibeVoice TTS
ComfyUI Node

VibeVoice TTS

A ComfyUI node in audio/tts with 16 inputs and 1 output.

By wildminder·Created 11 months ago·Updated 10 months ago· 589
VibeVoice TTS
  • speaker_1_voice
  • speaker_2_voice
  • speaker_3_voice
  • speaker_4_voice
  • AUDIO
model_name
text[1] Hello, this is a cloned voice. [2] And this is a generated voice, how cool is that?
quantize_llm_4bitfalse
attention_modesdpa
cfg_scale1.30
inference_steps10
seed42
do_sampletrue
temperature0.95
top_p0.95
top_k0
force_offloadfalse
Categoryaudio/tts

Inputs (16)

NameTypeDefaultDescription
model_nameCOMBOSelect the VibeVoice model to use. Official models will be downloaded automatically.
textSTRING[1] Hello, this is a cloned voice. [2] And this is a generated voice, how cool is that?The script for generation. Use '[1]' or 'Speaker 1:' for speakers. If a speaker in the script lacks a reference voice, it will be generated via zero-shot TTS.
quantize_llm_4bitBOOLEANfalseQuantize the Qwen2.5 LLM to 4-bit NF4 via bitsandbytes. Diffusion head stays BF16/FP32.
attention_modeCOMBOsdpaAttention implementation: Eager (safest), SDPA (balanced), Flash Attention 2 (fastest), Sage (quantized)
cfg_scaleFLOAT1.300.1–50Classifier-Free Guidance scale. Higher values increase adherence to the voice prompt but may reduce naturalness. Recommended: 1.3
inference_stepsINT101–500Number of diffusion steps for audio generation. More steps can improve quality but take longer. Recommended: 10
seedINT420–18446744073709550000Seed for reproducibility. Set to 0 for a random seed on each run.
do_sampleBOOLEANtrueEnable to use sampling methods (like temperature and top_p) for more varied output. Disable for deterministic (greedy) decoding.
temperatureFLOAT0.950–2Controls randomness. Higher values make the output more random and creative, while lower values make it more focused and deterministic. Active only if 'do_sample' is enabled.
top_pFLOAT0.950–1Nucleus sampling (Top-P). The model samples from the smallest set of tokens whose cumulative probability exceeds this value. Active only if 'do_sample' is enabled.
top_kINT00–500Top-K sampling. Restricts sampling to the K most likely next tokens. Set to 0 to disable. Active only if 'do_sample' is enabled.
force_offloadBOOLEANfalseForce model to be offloaded from VRAM after generation. Useful to free up memory between generations but may slow down subsequent runs.
speaker_1_voiceoptAUDIOReference audio for 'Speaker 1' or '[1]' in the script.
speaker_2_voiceoptAUDIOReference audio for 'Speaker 2' or '[2]' in the script.
speaker_3_voiceoptAUDIOReference audio for 'Speaker 3' or '[3]' in the script.
speaker_4_voiceoptAUDIOReference audio for 'Speaker 4' or '[4]' in the script.

Outputs (1)

NameTypeDescription
AUDIOAUDIO