InfiniteTalk Native Sampler for ComfyUI

A ComfyUI custom node package for generating lip-sync videos from audio using WAN 2.1 models with automatic frame extension for long audio.

Features

• 🎵 Automatic Frame Extension — Seamlessly handles long audio by automatically extending video generation using motion frames
• 🗣️ Lip-Sync Generation — Produces realistic talking head videos synchronized to input audio
• 👥 Multi-Speaker Support — Supports two speakers with separate masks for advanced scenarios
• 🎛️ Two Sampler Variants — Choose between standard KSampler or advanced custom sampler interfaces
• 🎞️ Configurable Framerate — Set output framerate (default 25fps) to match your target video format
• 🎬 Motion Frame Conditioning — Uses motion frames to maintain temporal consistency across extended generations

Nodes

1. InfiniteTalk Auto Sampler

Standard KSampler interface with familiar parameters:

• seed, steps, cfg, sampler_name, scheduler, denoise

Inputs: | Name | Type | Description | | ------------------------ | -------------------- | ------------------------------------------- | | model | MODEL | Base WAN 2.1 model | | model_patch | MODEL_PATCH | InfiniteTalk model patch | | positive | CONDITIONING | Positive conditioning | | negative | CONDITIONING | Negative conditioning | | vae | VAE | VAE for encoding/decoding | | audio_encoder_output_1 | AUDIO_ENCODER_OUTPUT | Primary audio encoder features | | audio | AUDIO | Input audio | | width | INT | Video width (default: 832) | | height | INT | Video height (default: 480) | | length | INT | Frames per pass (default: 81) | | motion_frame_count | INT | Motion frames for conditioning (default: 9) | | audio_scale | FLOAT | Audio conditioning scale (default: 1.0) | | framerate | INT | Output framerate in fps (default: 25) | | start_image | IMAGE | Optional starting image | | clip_vision_output | CLIP_VISION_OUTPUT | Optional CLIP vision conditioning | | audio_encoder_output_2 | AUDIO_ENCODER_OUTPUT | Optional second speaker audio | | mask_1, mask_2 | MASK | Speaker reference masks |

2. InfiniteTalk Auto Sampler (Advanced)

Custom sampler interface accepting NOISE/SAMPLER/SIGMAS objects for advanced workflows:

• Use with custom noise generators, samplers, and sigma schedules

Additional Inputs: | Name | Type | Description | | --------- | ------- | --------------------- | | noise | NOISE | Custom noise object | | sampler | SAMPLER | Custom sampler object | | sigmas | SIGMAS | Custom sigma schedule |

Installation

1. Clone this repository into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/vjumpkung/comfyui-infinitalk-native-sampler.git

2. Restart ComfyUI — the nodes will be automatically loaded.

or

Install from ComfyUI Manager search infinitetalk-native-sampler

Requirements

• ComfyUI with WAN 2.1 model support
• Compatible audio encoder (e.g., wav2vec2-base)
• comfy.ldm.wan.model_multitalk module (included in recent ComfyUI versions)

Note: A startup patch for the wav2vec2 fp16 dtype bug is included and applied automatically.

Workflow

A sample workflow is included in workflows/infinitetalk_native_loop_sampler_by_vjumpkung.json.

Basic Usage

1. Load your WAN 2.1 model and InfiniteTalk model patch
2. Encode your audio using a compatible audio encoder (wav2vec2-base recommended)
3. Connect the audio encoder output and audio to the InfiniteTalk sampler
4. Configure dimensions (width, height, length)
5. Set motion_frame_count for temporal consistency (default: 9)
6. Run the sampler — it will automatically calculate and execute multiple passes for long audio

Multi-Speaker Setup

1. Provide two audio encoder outputs (audio_encoder_output_1 and audio_encoder_output_2)
2. Provide corresponding reference masks (mask_1 and mask_2)
3. The system will combine audio features using "add" mode

How It Works

Frame Extension Algorithm

The sampler automatically handles long audio through multiple passes:

1. Base Pass: Generates initial length frames from motion frame (start image or zeros)
2. Extend Passes: Chains generation using last motion_frame_count frames as motion conditioning

Pass count calculation:

extend_frames = length - motion_frame_count
total_passes = 1 + ceil((total_frames - length) / extend_frames)

Audio Processing

• Linear interpolation from 50fps to the configured framerate
• Audio encoder output validated against model patch dimensions
• Compatible encoders: wav2vec2-base (blocks=12, channels=768)

Model Patching

The sampler applies patches to the model for each pass:

• attn2_patch: Cross-attention patch for audio conditioning
• attn1_patch: Self-attention patch for reference masks (when provided)
• InfiniteTalkOuterSampleWrapper: Handles motion frame conditioning

Parameters Guide

| Parameter | Default | Range | Description | | -------------------- | ------- | ----------------- | -------------------------------------------------- | | length | 81 | 1-4096 (step 4) | Frames generated per pass. Higher = more VRAM | | motion_frame_count | 9 | 1-33 | Frames used for motion conditioning between passes | | audio_scale | 1.0 | -10.0 to 10.0 | Strength of audio conditioning | | width | 832 | 16-4096 (step 16) | Output video width | | height | 480 | 16-4096 (step 16) | Output video height | | framerate | 25 | 1-120 | Output framerate; also controls audio resampling |

Tuning Tips

• Longer Audio: Reduce length or increase motion_frame_count to use fewer passes
• Better Consistency: Increase motion_frame_count for smoother transitions
• VRAM Management: Lower length if running out of memory
• Audio Strength: Adjust audio_scale to control lip-sync intensity

Technical Details

Latent Shape Convention

Shape format: [batch, channels, temporal, height//8, width//8]

• Channels: 16 (for WAN)
• Temporal: ((length - 1) // 4) + 1 (4x temporal compression)

Audio Feature Validation

Audio encoder dimensions must match:

• model_patch.model.audio_proj.blocks (e.g., 12)
• model_patch.model.audio_proj.channels (e.g., 768)

Troubleshooting

"Audio encoder output dimensions do not match"

Ensure you're using a compatible audio encoder (wav2vec2-base for standard models).

Out of Memory

Reduce length parameter to generate fewer frames per pass.

"Both mask_1 and mask_2 are required"

When using two audio encoders, both masks must be provided.

"Input type (float) and bias type (c10::Half) should be the same"

This is a ComfyUI core bug that occurs when the wav2vec2 audio encoder is loaded in fp16. This package automatically patches the issue at startup via _patches.py — no manual action needed. If the error persists, ensure you are using a recent version of this package.

License

• Some codebase are from ComfyUI

Credits

• Based on WAN 2.1 model architecture
• InfiniteTalk implementation for lip-sync generation

The entire codebase is generate from Claude Code.