LongCat Avatar FP8: New ComfyUI Workflow with Audio Sync

LongCat Avatar FP8: New ComfyUI Workflow with Audio Sync

LongCat Avatar on ComfyUI has just been updated with FP8 optimization. This guide walks you through setting up the workflow, configuring models, and generating talking videos with audio synchronization.

TL;DR

The updated LongCat Avatar FP8 workflow includes:

• FP8 converted model - smaller file, efficient VRAM usage
• Audio sync support - lip sync to any audio file
• Block swap options - run on 24GB or lower cards
• Five sampling stages - chained for extended generation
• Multi-GPU support - offload encoding to second GPU

Required Models and Files

Core Model Files

Additional Files

Since LongCat Avatar is based on the 14B Wan 2.1 model, you should already have the text encoder and VAE files if you've run Wan workflows before.

GGUF files for lower VRAM systems are expected to be available soon.

Workflow Configuration

Model Loading Section

FP8 Model Setup:

Main Model: longcat-avatar-fp8.safetensors
Block Swap: 20 (for 24GB) or 40 (for lower VRAM)

LoRA Configuration:

• Load Acceleration LoRA
• Load LongCat weight

VRAM Settings

Input Configuration

Image Input

Prepare your source image in the input section:

• Upload character photo
• Ensure good lighting and clear face
• Standard portrait orientation works best

Audio Configuration

Configure your audio file settings:

Audio File: your_audio.wav
Start Time: 0 (seconds)
End Time: [total length] (seconds)

The workflow will sync generated video to this audio clip.

Video Settings

Frame Configuration

Note: 77 frames appears to be the current maximum for stable generation per sampling run.

Multi-GPU Setup (Optional)

If you have multiple GPUs, you can offload the text encoder:

1. Enable the Multi-GPU Text Encoder node
2. Specify which GPU for encoding
3. Main model stays on primary GPU
4. Reduces VRAM pressure on main card

Disable this node if running single GPU setup.

Sampling Pipeline

The workflow includes five sampling stages chained together:

Stage 1 → Stage 2 → Stage 3 → Stage 4 → Stage 5

Each stage builds on the previous, enabling extended video generation.

Why Five Stages?

• Maintains quality over longer durations
• Each stage handles a portion of the timeline
• Chained approach prevents quality degradation
• Enables multi-minute output

Expected Results

What Works Well

• Facial expressions - natural and responsive
• Background movements - realistic motion
• Audio syncing - better than expected
• Overall coherence - good quality output

Known Limitations

• Hands and fingers - may look off
• Complex gestures - occasional artifacts
• Extended duration - quality gradually decreases

Troubleshooting

Node Errors

If you encounter errors:

1. Check VRAM allocation
2. Increase block swap value
3. Verify all model files are present
4. Restart ComfyUI and reload workflow

VRAM Issues

For out-of-memory errors:

1. Increase block swap (up to 40)
2. Close other GPU applications
3. Reduce resolution if possible
4. Consider cloud deployment

Audio Sync Problems

If audio doesn't sync properly:

1. Verify audio file format (WAV recommended)
2. Check start/end time settings
3. Ensure audio length matches intended video

Performance Expectations

Generation Time

Quality vs Speed

Higher block swap = slower but more stable Lower block swap = faster but more VRAM needed

Future Updates

GGUF Files

Low VRAM users can expect GGUF quantized files soon, enabling:

• Running on 12GB or lower
• Significantly reduced memory footprint
• Slightly reduced quality (acceptable tradeoff)

Workflow Improvements

The current workflow is work-in-progress. Expect:

• Better hand/finger generation
• Improved expression range
• More configuration options

Related Resources

1. LongCat Avatar ComfyUI Tutorial - Complete avatar setup guide

2. Build Your Own AI Avatar with LongCat - Avatar creation workflow

3. LongCat FP8 Workflow: Memory-Efficient Video - FP8 optimization details

4. ComfyUI LongCat Workflow Low VRAM - Techniques for limited hardware

LongCat Avatar FP8 brings efficient, audio-synced talking video generation to ComfyUI with optimizations for 24GB cards and multi-GPU support.