LongCat-Video-Avatar 1.6 Technical Report

Meituan LongCat Team, Xunliang Cai, Yiqi Chen, Feng Gao, Xiaoyu Kong, Le Li, Hongyu Liu, Xiaoming Wei, Tianyu Yang, Yong Zhang, Sijie Zhao

TL;DR

LongCat-Video-Avatar 1.6 is an open-source framework that significantly advances audio-driven avatar generation by focusing on two persistent "boss battles" in the field: natural expressive motion and long-term temporal consistency. By combining reinforcement learning (GRPO) with on-policy self-distillation, the team has produced a model capable of generating high-fidelity videos exceeding 10 minutes that outperform major commercial competitors like HeyGen and Kling.

What Problem Does This Paper Address?

While AI avatars have become visually impressive, they often fall into the "uncanny valley" for two reasons. First, unnatural articulation: many models achieve perfect lip-sync but with exaggerated, "strained" mouth movements that don't match human speech dynamics. Second, autoregressive drift: in long videos, small errors in color or texture accumulate. A common artifact is the "reddening lip" syndrome, where an avatar’s lips progressively become over-saturated and distorted over several minutes of footage.

The Meituan LongCat team tackles these issues by moving beyond simple clip-level training, focusing instead on how a model behaves when it has to "roll out" a long sequence based on its own previously generated frames.

Methods

The authors introduce a dual-branch post-training pipeline that optimizes a frozen base model using lightweight LoRA adapters.

  1. Reinforcement Learning Branch (GRPO): To improve facial naturalness, they use Group Relative Policy Optimization (GRPO). Unlike standard RL, GRPO doesn't require a separate "value network" to estimate rewards. Instead, it generates a group of candidate videos for the same input and ranks them against each other. This allows the model to learn "relaxed" and speech-appropriate mouth movements based on perceptual rewards.
  2. Distillation & Long-Video Branch:
    • Few-step Distillation (DMD2): They distilled the original 50-step diffusion process down to just 8 steps, drastically increasing inference speed without losing quality.
  3. On-Policy Self-Distillation (OPSD): This is the secret sauce for long videos. During training, the "student" model generates a sequence of clips. Each new clip is conditioned on the student's own (imperfect) previous output. An "EMA Teacher" model provides guidance, teaching the student how to correct its own drift and maintain visual consistency over time.
  4. Revised CFG Path: They discovered that standard Classifier-Free Guidance (CFG) often caused the "red lip" artifact. They implemented a revised guidance path ($00 \rightarrow 0T \rightarrow A0$) that balances text and audio conditioning to maintain natural skin tones and facial details.

Datasets

The model was trained using the extensive data resources from version 1.5, which includes:

  • Source Pool: Close-up facial videos, interviews, acted performances, and music videos.
  • Annotations: Detailed metadata on subject geometry, speaking activity, and emotional expression.
  • Evaluation: The primary benchmark used was EvalTalker, consisting of 508 image-speech pairs.
  • Arena: A "Human Preference Arena" was conducted using 33,051 valid pairwise votes to rank the model against commercial SOTA (State-of-the-Art) systems.

Key Insights and Ideas

  • The "History" Problem: The authors realized that optimizing for single, clean clips isn't enough for real-world use. Models must be trained on their own "imperfect histories" to learn how to stay stable during 10+ minute generations.
  • Detail vs. Fidelity: A fascinating finding in the Arena results showed that while OmniHuman-2.0 ranked first, it often "enhanced" the reference image (adding detail that wasn't there). LongCat 1.6 focuses on identity preservation, staying truer to the original source image even if it means looking "softer" than models that hallucinate extra detail.
  • Guidance Matters: The discovery that specific CFG paths ($00 \rightarrow 0T \rightarrow A0$) could solve color saturation issues suggests that many "visual bugs" in diffusion models are actually steering problems rather than fundamental architecture flaws.

Key Results

  • Arena Ranking: LongCat-Video-Avatar 1.6 ranked 2nd overall with an Elo score of 1068.1. It outperformed HeyGen (908.1), OmniHuman-1.5 (880.6), and Kling Avatar 2.0 (854.7).
  • Human-Likeness: The model saw a significant jump in human-likeness scores, moving from 3.006 (v1.5) to 3.322 (v1.6).
  • Consistency: The "aggregate consistency issue rate" (a measure of visual drift) dropped from 19.9% to 11.6%.
  • Efficiency: By reducing denoising steps from 50 to 8, the model is now practical for high-throughput production environments.

Limitations and Future Work

The authors noted a slight performance dip in animated/stylized inputs (dropping from 2.246 to 2.114). They attribute this to the fact that their long-video training set (OPSD) was primarily composed of realistic human videos. Future work will likely involve diversifying the long-form training data to include more diverse artistic styles. Additionally, while the model handles 10-minute videos well, the team aims to push the boundaries of "infinite" generation while maintaining even tighter identity anchors.

Summary updated

Abstract

Recent advances in audio-driven avatar video generation have enabled increasingly realistic synthesis, yet achieving precise lip-audio synchronization, natural, human-like motion, and visual consistency over long sequences remains challenging. We present LongCat-Video-Avatar 1.6, which builds on LongCat-Video-Avatar 1.5 to improve expressive naturalness and reduce error accumulation during extended generation, e.g., exaggerated lip movements and overly red, temporally unstable lip coloration, enhance the naturalness of facial expressions and coordinated head and body movements. These improvements are driven by a refined post-training pipeline that integrates reinforcement learning, improved classifier-free guidance strategies for DMD-based few-step distillation, long-video rollouts, and on-policy distillation. Together, these refinements improve expressive realism and enable the generation of videos exceeding 10 minutes while maintaining visual quality and temporal consistency. Arena evaluations show that LongCat-Video-Avatar 1.6 outperforms its predecessor and competing commercial audio-driven avatar models, demonstrating the effectiveness of the improved post-training pipeline.
BibTeX
@misc{team2026longcat,
  author = {Meituan LongCat Team and Xunliang Cai and Yiqi Chen and Feng Gao and Xiaoyu Kong and Le Li and Hongyu Liu and Xiaoming Wei and Tianyu Yang and Yong Zhang and Sijie Zhao},
  title = {{LongCat-Video-Avatar 1.6 Technical Report}},
  year = {2026},
  eprint = {2611.18822},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18822},
}
Figure