LongCat-Video-Avatar

LongCat-Video-Avatar

Meituan's open audio-driven avatar video model that stays stable past ten minutes

AIOfficial site

Description

Two things ruin AI talking-head videos: lips that drift off the audio or look strained, and long clips where the face slowly turns into someone else.

LongCat-Video-Avatar is Meituan LongCat's open-source audio-driven human video generation model, built on the LongCat-Video base model. Give it audio plus a text prompt or a reference image and it produces a video of the person speaking, for single or multiple characters, and it can also continue an existing video.

What's new in 1.6

More natural lips: fewer exaggerated or strained mouth movements, less overly red lip color and less flicker around the mouth.
Better-coordinated motion across face, head and body.
Stable long videos: less error accumulation during autoregressive continuation, with videos over 10 minutes keeping visual quality and identity, according to the team.
Faster inference from end-to-end pipeline and execution optimizations.

In a human preference study against 1.5, OmniHuman-2.0, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0, with 33,051 valid votes, 1.6 scored a Bradley–Terry Elo of 1068.1, second of six behind OmniHuman-2.0.

Running it

The code lives in the LongCat-Video GitHub repository and needs Python 3.10, PyTorch and FlashAttention, plus both the LongCat-Video and LongCat-Video-Avatar-1.6 weights from Hugging Face. The checkpoint includes an optional INT8 model and FP8 is supported, one quantization mode at a time. On consumer GPUs, the community route is ComfyUI: Kijai's WanVideoWrapper already supports version 1.5, with fp8 and GGUF weights available.

Released under the MIT license, it suits talking-head content, virtual presenters and course videos that need long on-camera time.