ComfyUI_iAskAsk_SOULX_FlashHead

ComfyUI custom nodes for SoulX-FlashHead — generate talking head videos from a reference image and driving audio.

Provider Website: http://iaskask.com/

Features

• Realistic talking head video generation from a single portrait photo + audio
• Supports both Pro and Lite model variants
• Long audio slicing support — generates coherent videos for arbitrary-length audio
• Custom checkpoint paths for model flexibility
• BFloat16/Float16 autocast for memory efficiency

Requirements

Install via ComfyUI manager or manually:

pip install -r requirements.txt

Key dependencies: optimum-quanto, diffusers, transformers, librosa, imageio-ffmpeg, av, mediapipe, loguru.

Setup

ComfyUI/custom_nodes/
└── ComfyUI_iAskAsk_SoulX_FlashHead/

This plugin uses the flash_head module from SoulX-FlashHead (bundled as a submodule).

Model Download

Two models are required: SoulX-FlashHead-1_3B (~12GB) and wav2vec2-base-960h. Two download scripts are provided:

Region	Script	Source
Global	python download_model_huggingface.py	HuggingFace Hub
China	python download_model_modelscope.py	ModelScope (falls back to HuggingFace for wav2vec2)

Models are saved to models/ by default. After downloading, the Loader node's default paths will work automatically:

• ComfyUI/models/Soul-AILab/SoulX-FlashHead-1_3B/
• ComfyUI/models/facebook/wav2vec2-base-960h/

Nodes

iAskAsk Local SoulX FlashHead Loader

Loads the FlashHead pipeline.

Parameter	Type	Default	Description
model_type	pro / lite	lite	Model variant
ckpt_dir	STRING	ComfyUI/models/Soul-AILab/SoulX-FlashHead-1_3B	Custom checkpoint path
wav2vec_dir	STRING	ComfyUI/models/facebook/wav2vec2-base-960h	Custom wav2vec path

iAskAsk Local FlashHead Sampler

Generates video from pipeline, reference image, and audio.

Parameter	Type	Default	Description
pipeline	Pipeline	—	From the Loader node
ref_audio	AUDIO	—	Driving audio signal
avatar_image	IMAGE	—	Reference face image
seed	INT	42	Random seed (use -1 for random)
width	INT	512	Output video width
height	INT	512	Output video height

Usage

1. Install SoulX-FlashHead (submodule) per its setup instructions.
2. Download models using one of the scripts above.
3. Load the workflow from workflow/ (or create your own):
   • iAskAsk Local SoulX FlashHead Loader → configure model paths
   • Load Image → your reference portrait
   • Load Audio → driving audio
   • iAskAsk Local FlashHead Sampler → connect pipeline, image, audio
   • Preview/Export Video → view result

Example Assets

Sample files are included in assets/ for quick testing:

• girl_001.png — reference portrait
• girl_001_voice.wav — driving audio
• girl_001_out.mp4 — sample output

Testing Notes

Manual testing by the author:

• Launch ComfyUI, fix width & height to 512×512 (recommended)
• The test can run continuously for 1+ minutes of generation without issues

Acknowledgements

• SoulX-FlashHead — core model and inference pipeline
• iAskAsk — ComfyUI integration provider

Development

This project was built completely using opencode — an AI coding tool — with the DeepSeek V4 Flash Free model (truly free, via opencode zen).
Developer contributed debugging assistance during development.