Meituan Open-Sources LongCat-Video-Avatar1.5, Outperforming Mainstream Closed Models

Meituan's LongCat large model team has officially open-sourced LongCat-Video-Avatar 1.5, a commercial-grade digital human video generation model. This version represents a full shift from open-source state-of-the-art to real-world commercial deployment, with major enhancements in lip sync, physical realism, long-video stability, multi-person interactions, and efficient inference.

Three Key Upgrades to Overcome Commercialization Hurdles

To make digital humans work reliably across various real-world applications, LongCat-Video-Avatar 1.5 tackles persistent problems like jitter, distortion, and high latency in traditional digital human videos through three holistic improvements:

Commercial-Grade Audio Encoding Upgrade

The model’s audio feature extraction encoder has been upgraded from Wav2Vec2 to Whisper-large. With more parameters and richer multilingual prior knowledge, it now captures subtle phoneme variations and speech rhythm. This improves lip synchronization for complex audio such as long sentences, fast speech, and singing, while also enabling natural coordination between facial expressions, head movements, and speech—significantly reducing frame skipping and identity drift in longer videos.

Robust Open-Domain Generalization via Multi-Stage Data Enhancement

To ensure stable performance across diverse subjects—real people, virtual idols, anime characters, and animals—the team developed a multi-stage data pipeline incorporating offline annotation and online validation, and introduced three targeted types of enhanced data:

Multi-Person Data: Using active speaker detection, it removes audio-visual ambiguity in multi-person scenes, accurately differentiating speakers from listeners.

Quiet Data: By selecting videos with no speech, the model learns natural micro-expressions during silent states, preventing unwanted mouth movements in non-speaking characters.

Emotion Data: Combined with frame-level emotion recognition filtering, it incorporates emotional variations, helping the model grasp the deep link between speech and facial expressions.

Hand Alignment and Temporal Continuity with GRPO

For use cases like e-commerce live streams and product demos where hands are frequently visible, the model introduces GRPO (Human Preference Alignment), which refines reward signals to the frame level and adds a first-frame hand detection mechanism. This substantially reduces common issues like hand distortion, local structural collapse, and inconsistent motion.

15× Faster Inference: Eliminating Costly Compute Requirements

Cost is another critical factor for commercial deployment. LongCat-Video-Avatar 1.5 uses DMD (Distributed Matching Distillation) technology, compressing the original 50-step generation process to just 8 steps. In addition, the team replaced the conventional three-model parallel setup with a single shared base model plus multiple LoRA adapters, dramatically reducing VRAM usage.

In real-world tests, the model delivers roughly 15× faster inference, generating a 10-second video in about one minute.

Benchmark Evaluation: Outperforming Top Industry Models

Using the EvalTalker benchmark, 770 evaluators and 10 domain experts performed structured quality analysis on videos from complex domains including news, education, and entertainment. The results show that LongCat-Video-Avatar 1.5 excels across several key metrics:

User Preference Win Rate: It beats Kling Avatar2.0 by 65.9%, OmniHuman-1.5 by 61.1%, and HeyGen by 54.3%.

Single and Multi-Person Scenario Scores: The single-person score reaches 3.336, well above competitors like HeyGen; the multi-person score is 2.730, significantly outperforming InfiniteTalk (2.339).

Video Stability: The subject deformation rate is only 23.1%, and background deformation rate is just 9.4%;the frame skipping rate is as low as 0.8%, the best among all compared models.

Audio-Visual Coordination: The face-body sync issue rate fell to 5.1%, and the lip sync issue rate dropped to 29.8%, both surpassing traditional commercial systems.

The Meituan LongCat large model team says that open-sourcing LongCat-Video-Avatar 1.5 is more than a version update—it's an invitation to the global developer and creator community. They hope this model will serve as a verifiable, improvable technical foundation, helping to push the real-world boundaries of digital human video applications.

Open Source Links:

Github: https://github.com/meituan-longcat/LongCat-Video

HuggingFace: https://huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.5

Tech Report: https://github.com/meituan-longcat/LongCat-Video/blob/main/assets/LongCat-Video-Avatar-1.5-Tech-Report.pdf

Project Page: https://meigen-ai.github.io/LongCat-Video-Avatar-1.5-Page/

Modelscope: https://www.modelscope.cn/models/meituan-longcat/LongCat-Video-Avatar-1.5/summary