Live Conversation — Real-time voice chat with AI using your microphone
Batch Inference — Process audio files through the PersonaPlex model
Voice Presets — 18 different voice options (natural and variety styles)
Zero-Shot Voice Cloning — Clone any voice from a 3-10 second audio sample
Quantization Support — Run on 8-16GB VRAM GPUs with 8-bit or 4-bit quantization
Persona Control — Customize AI behavior with text prompts
Full Settings Control — Adjust sampling, temperature, and timing parameters
Smart Server Startup — Browser opens only after server is fully ready
Server Cleanup — Automatic cleanup on exit, plus manual stop node

Screenshots

PersonaPlex Conversation Server Node

Start the live conversation server with model selection and options.

PersonaPlex Settings Node (Optional)

Configure all sampling, temperature, and timing parameters in one place.

Web UI - Live Conversation

Real-time voice conversation with customizable personas and voice selection.

Installation

1. Clone or Download

cd ComfyUI/custom_nodes
git clone https://github.com/squarewulf/ComfyUI_PersonaPlexMF.git

2. Install Dependencies

Important: The sphn library must be version 0.1.x (version 0.2+ has breaking API changes).

cd ComfyUI/custom_nodes/ComfyUI_PersonaPlexMF

# Install requirements (includes pinned sphn version)
pip install -r requirements.txt

# Install bundled moshi package (required for PersonaPlex features)
pip install -e personaplex_src/moshi

For Windows portable ComfyUI, use the embedded Python:

# From ComfyUI_windows_portable directory
.\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI_PersonaPlexMF\requirements.txt
.\python_embeded\python.exe -m pip install -e ComfyUI\custom_nodes\ComfyUI_PersonaPlexMF\personaplex_src\moshi

3. Download Models

Download from HuggingFace nvidia/personaplex-7b-v1 and place in ComfyUI/models/personaplex/:

File	Description
`model.safetensors`	Moshi 7B LM weights (~14GB)
`tokenizer-e351c8d8-checkpoint125.safetensors`	Mimi encoder/decoder
`tokenizer_spm_32k_3.model`	Text tokenizer
`voices.tgz`	Voice presets (extract to `voices/` folder)

The web UI for live conversation is included in this repo (no separate download needed).

Nodes

PersonaPlex Conversation Server

Starts a WebSocket server with a web UI for live voice conversation.

Inputs:

Input	Type	Description
`moshi_model`	dropdown	Moshi LM weights file
`mimi_model`	dropdown	Mimi encoder/decoder file
`tokenizer`	dropdown	Text tokenizer file
`port`	int	Server port (default: 8998)
`device`	dropdown	cuda or cpu
`cpu_offload`	bool	Enable for low VRAM GPUs
`quantize`	dropdown	Quantization level (see Model Loader)
`open_browser`	bool	Auto-open browser (waits for server ready)
`settings`	optional	Connect PersonaPlex Settings node

Output: server_url — URL to access the conversation UI

PersonaPlex Model Loader

Loads PersonaPlex models for batch inference.

Inputs:

Input	Type	Description
`moshi_model`	dropdown	Moshi LM weights file
`mimi_model`	dropdown	Mimi encoder/decoder file
`tokenizer`	dropdown	Text tokenizer file
`device`	dropdown	cuda or cpu
`cpu_offload`	bool	Enable for low VRAM GPUs
`quantize`	dropdown	Quantization level (see below)

Output: personaplex_model — Model bundle for inference

Quantization Options:

Option	VRAM Required	Quality	Notes
`none`	~14GB	Best	Full precision (default)
`8bit`	~8-10GB	Good	8-bit quantization
`4bit`	~5-7GB	Acceptable	4-bit NF4 quantization

Quantization:

Quantization works out of the box - no additional packages needed! Uses PyTorch native dynamic quantization.

VRAM Requirements:

none (full precision): ~14GB VRAM
8bit: ~8-10GB VRAM
4bit: ~5-7GB VRAM

Important Notes:

Works on Windows and Linux with zero dependencies
Quantization and cpu_offload cannot be used together
If you have 24GB+ VRAM (e.g., RTX 3090/4090), use none for best quality

PersonaPlex Inference

Runs speech-to-speech inference on audio input.

Inputs:

Input	Type	Description
`personaplex_model`	model	From Model Loader
`audio`	AUDIO	Input audio
`settings`	optional	Connect PersonaPlex Settings node
`voice_audio`	AUDIO	Optional: Audio sample for voice cloning (3-10s)
`voice_preset`	dropdown	Voice to use (ignored if voice_audio connected)
`text_prompt`	string	Persona/behavior prompt

Outputs:

audio — Generated speech response
text_output — Transcription of response

Voice Cloning: To clone any voice, connect an audio sample (3-10 seconds of clear speech) to the voice_audio input. This overrides the voice preset and generates responses in the cloned voice.

PersonaPlex Settings

Configure all inference parameters. Connect to Inference or Server nodes.

Sampling Settings:

Setting	Default	Description
`temp_audio`	0.8	Audio sampling temperature
`temp_text`	0.7	Text sampling temperature
`topk_audio`	250	Top-K for audio tokens
`topk_text`	25	Top-K for text tokens
`use_sampling`	true	Use sampling vs greedy

Timing Settings:

Setting	Default	Description
`silence_duration`	0.5	Seconds before response starts

Other Settings:

Setting	Default	Description
`seed`	-1	Random seed (-1 = random)
`voice_preset`	default	Voice override
`text_prompt`	""	Persona prompt override
`host`	0.0.0.0	Server bind address
`use_ssl`	false	Enable HTTPS
`gradio_tunnel`	false	Enable remote access

PersonaPlex Stop Server

Stops the running conversation server. Useful when you need to free up GPU memory or restart with different settings.

Output: status — Result of the stop operation

Voice Presets

Category	Voices
Natural Female	NATF0, NATF1, NATF2, NATF3
Natural Male	NATM0, NATM1, NATM2, NATM3
Variety Female	VARF0, VARF1, VARF2, VARF3, VARF4
Variety Male	VARM0, VARM1, VARM2, VARM3, VARM4

Example Prompts

Assistant:

You are a wise and friendly teacher. Answer questions or provide advice in a clear and engaging way.

Customer Service:

You work for AeroRentals Pro which is a drone rental company and your name is Alex.

Casual Conversation:

You enjoy having a good conversation. Have a casual discussion about technology.

Hardware Requirements

Configuration	VRAM	Notes
Recommended	24GB+	Full speed inference
Minimum	12GB	Use `cpu_offload=True` (slower)

Troubleshooting

"Torch not compiled with CUDA"

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

"No module named 'sphn'" or "OpusStreamReader has no attribute 'read_pcm'"

The sphn library version 0.2+ has breaking API changes. Install version 0.1.x:

pip install "sphn>=0.1.4,<0.2"

For Windows portable ComfyUI:

.\python_embeded\python.exe -m pip install "sphn>=0.1.4,<0.2"

"unrecognized arguments: --voice-prompt-dir"

A different moshi package is installed that conflicts with PersonaPlex. Fix by installing the bundled version:

cd ComfyUI/custom_nodes/ComfyUI_PersonaPlexMF
pip install -e personaplex_src/moshi

Server won't start

Check ComfyUI console for error messages
Ensure all model files are in ComfyUI/models/personaplex/
On Windows, make sure paths don't contain special characters

Port already in use

Change the port number in the node settings
Or use the "PersonaPlex Stop Server" node to stop the previous instance

WebSocket disconnects immediately

Check the ComfyUI console for [PersonaPlex Server] error messages
Ensure sphn version is < 0.2 (see above)
Wait for "Server is ready!" message before connecting

Quantization

Quantization uses PyTorch native and should work automatically. At startup you'll see:

[PersonaPlex] Quantization available (backend: pytorch_native)

Note: If you have 24GB+ VRAM (RTX 3090/4090), you don't need quantization - use none for best quality.

Changelog

v1.3.0

Added 8-bit and 4-bit quantization support for low VRAM GPUs
New quantize option in both Model Loader and Conversation Server
Run on GPUs with 8-16GB VRAM instead of 24GB+
Graceful fallback to full precision if bitsandbytes fails
Improved Windows compatibility with better error messages
Requires bitsandbytes (optional, Linux recommended)

v1.2.0

Added zero-shot voice cloning from any audio sample
New voice_audio input on PersonaPlexInference node
Connect any audio (3-10 seconds recommended) to clone that voice

v1.1.0

Fixed server startup and WebSocket connection stability
Browser now waits until server is fully ready
Pinned sphn version to < 0.2 (required for audio streaming)
Improved error messages and server cleanup

v1.0.0

Initial release

License

PersonaPlex model weights: NVIDIA Open Model License
This ComfyUI wrapper: MIT License

Author

squarewulf