ComfyUI Node
VibeVoice Multiple Speakers
Generate multi-speaker conversations with up to 4 distinct voices using Microsoft VibeVoice
VibeVoice Multiple Speakers
- speaker1_voice
- speaker2_voice
- speaker3_voice
- speaker4_voice
- lora
- audio
◄text[1]: Hello, this is the first speaker.
[2]: Hi there, I'm the second speaker.
[1]: Nice to meet you!
[2]: Nice to meet you too!►
◄modelNo models found►
◄attention_typeauto►
◄quantize_llmfull precision►
◄free_memory_after_generatetrue►
◄diffusion_steps20►
◄seed42►
◄cfg_scale1.30►
◄use_samplingfalse►
◄temperature0.95►
◄top_p0.95►
◄voice_speed_factor1.00►
CategoryVibeVoiceWrapper
Inputs (17)
| Name | Type | Default | Description |
|---|---|---|---|
| text | STRING | [1]: Hello, this is the first speaker. [2]: Hi there, I'm the second speaker. [1]: Nice to meet you! [2]: Nice to meet you too! | Text with speaker labels. Use '[N]:' format where N is 1-4. Gets disabled when connected to another node. |
| model | COMBO | No models found | Select a model from ComfyUI/models/vibevoice/ folder. Large is recommended for multi-speaker |
| attention_type | COMBO | auto | Attention implementation. Auto selects the best available, eager is standard, sdpa is optimized PyTorch, flash_attention_2 requires compatible GPU, sage uses quantized attention for speedup (CUDA only) |
| quantize_llm | COMBO | full precision | Dynamically quantize only the LLM component for non-quantized models. 4bit: major VRAM savings with minimal quality loss. 8bit: good balance of quality and memory usage. Full precision: original quality. Note: ignored for pre-quantized models. Requires CUDA GPU. |
| free_memory_after_generate | BOOLEAN | true | Free model from memory after generation to save VRAM/RAM. Disable to keep model loaded for faster subsequent generations |
| diffusion_steps | INT | 201–100 | Number of denoising steps. More steps = theoretically better quality but slower. Default: 20 |
| seed | INT | 420–4294967295 | Random seed for generation. Default 42 is used in official examples |
| cfg_scale | FLOAT | 1.300.5–3.5 | Classifier-free guidance scale (official default: 1.3) |
| use_sampling | BOOLEAN | false | Enable sampling mode. When False (default), uses deterministic generation like official examples |
| speaker1_voiceopt | AUDIO | Optional: Voice sample for Speaker 1. If not provided, synthetic voice will be used. | |
| speaker2_voiceopt | AUDIO | Optional: Voice sample for Speaker 2. If not provided, synthetic voice will be used. | |
| speaker3_voiceopt | AUDIO | Optional: Voice sample for Speaker 3. If not provided, synthetic voice will be used. | |
| speaker4_voiceopt | AUDIO | Optional: Voice sample for Speaker 4. If not provided, synthetic voice will be used. | |
| loraopt | LORA_CONFIG | Optional: LoRA configuration from VibeVoice LoRA node | |
| temperatureopt | FLOAT | 0.950.1–2 | Only used when sampling is enabled |
| top_popt | FLOAT | 0.950.1–1 | Only used when sampling is enabled |
| voice_speed_factoropt | FLOAT | 1.000.8–1.2 | 1.0 = normal speed, <1.0 = slower speed, >1.0 = faster speed (applies to all speakers) |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |