LongCat Video Avatar 1.5

The Meituan team has released LongCat-Video-Avatar 1.5, an open-source framework for audio-driven human video generation. It is an upgraded version focused on commercial-grade stability and extreme empirical optimization.

Key Improvements

• Whisper-Large Audio Encoder: Replaces the previous Wav2Vec2 to achieve much smoother and more natural lip synchronization • Commercial Stability: Supports long video generation while maintaining precise lip-sync, full-body temporal stability, and strict identity consistency • Stylized Domain Generalization: Generalizes robustly to complex real-world conditions including animation, animals, multi-person interactions, and object manipulation • Efficient 8-Step Inference: DMD2-based step distillation accelerates inference to 8 NFE, balancing cost-efficient serving with excellent visual fidelity

Supported Tasks: Natively supports Audio-Text-to-Video (AT2V), Audio-Text-Image-to-Video (ATI2V), and Video Continuation, and is compatible with both single-stream and multi-stream audio input.

A human evaluation benchmark was introduced, comprising a total of 508 image-audio source pairs across 6 application scenarios (news broadcast, knowledge education, daily life, entertainment, singing, commercial promotion), 2 languages (Chinese/English), and 2 visual styles (realistic/animated). 770 crowdsourced evaluators rated human-likeness on a 1-5 scale, collecting a total of 13,240 judgments, and 10 domain experts conducted structured quality analysis across 4 dimensions.