LongCat Video Avatar 1.5: Meituan's Open Audio-Driven Avatar Model

LongCat-Video-Avatar 1.5 is Meituan's upgraded open-source framework for audio-driven human video generation. Built on the LongCat-Video foundation model, it produces stable, production-grade avatar videos with native support for Audio-Text-to-Video (AT2V), Audio-Text-Image-to-Video (ATI2V) and video continuation, working with single or multiple audio streams.

What's New in 1.5

• Whisper-Large audio encoder replaces Wav2Vec2, delivering noticeably smoother and more natural lip dynamics
• 8-step distilled inference based on DMD2, balancing cost-effective serving with high visual fidelity
• Production-ready stability — accurate lip-sync, full-body temporal stability and strict identity consistency over long videos
• Stylized domain generalization — robust across anime, animals, multi-person interactions and object handling

The project page includes side-by-side comparisons against commercial services for lip sync and 3D animation:

ComfyUI Support

The community node pack ComfyUI-LongCat-Avatar adapts the official Avatar 1.5 pipeline to ComfyUI: nine nodes covering single- and multi-audio avatar generation, Whisper-large-v3 audio conditioning, the required DMD distill LoRA, and three weight modes — single-file .safetensors DiT, official sharded and official INT8 sharded checkpoints — at 480p and 720p.

Model files go into the standard ComfyUI folders (models/loras, models/vae, models/audio_encoders, models/diffusion_models), and the nodes can auto-download the official sharded checkpoints. The pack's example workflow:

Availability

Weights, the technical report and inference code are available through the official LongCat-Video repository and the project page.