Introduction

Meituan-LongCat has announced the release of LongCat-Video-Avatar 1.5, an upgraded open-source framework for audio-driven human video avatar generation. This new version, built upon the LongCat-Video foundation model, places a particular emphasis on empirical optimization and production-readiness for real-world scenarios.

The goal is to provide a robust and stable solution for commercial-grade avatar video synthesis, natively supporting tasks such as Audio-Text-to-Video (AT2V), Audio-Text-Image-to-Video (ATI2V), and Video Continuation. The framework is also compatible with both single-stream and multi-stream audio inputs, expanding its application versatility.

Technical Details and Key Features

LongCat-Video-Avatar 1.5 introduces several significant innovations. The audio encoder has been upgraded to Whisper-Large, replacing the previous Wav2Vec2. This change results in noticeably smoother and more natural lip dynamics, enhancing the overall visual experience.

The framework ensures production-ready stability, with accurate lip-synchronization, full-body temporal stability, and the ability to generate long videos while maintaining strict identity consistency. Furthermore, the model demonstrates remarkable generalization to stylized domains, effectively handling anime, animals, and complex real-world conditions, such as multi-person interactions and object handling. A crucial aspect for deployment is inference efficiency: thanks to DMD2-based step distillation, the process is accelerated to just 8 NFE (Number of Function Evaluations). This effectively balances serving costs with exceptional visual fidelity, a key factor for organizations evaluating the TCO of their AI infrastructures.

Human Evaluation and Implications

To validate its capabilities, LongCat-Video-Avatar 1.5 underwent a rigorous human evaluation benchmark specifically designed for audio-driven digital human generation. The benchmark covers six application scenarios (News Broadcasting, Knowledge Education, Daily Life, Entertainment, Singing, Commercial Promotion), two languages (Chinese/English), and two visual styles (Realistic/Animated), utilizing a total of 508 image-audio source pairs.

The evaluation methodology included a subjective track, with 770 crowdsourced evaluators providing 13,240 judgments on a 1-5 human-likeness scale, and an objective track, where 10 domain experts conducted structured quality analysis across four dimensions: Physical Rationality, Harmony (Audio-Visual Coordination), Temporal Stability, and Identity Consistency. This in-depth evaluation underscores the model's reliability and quality, fundamental aspects for companies seeking robust and verifiable AI solutions, especially in contexts where credibility and visual consistency are critical.

Prospects for On-Premise Deployment

While the source does not directly specify deployment scenarios, the features of LongCat-Video-Avatar 1.5 make it particularly appealing for on-premise and self-hosted strategies. As an open-source framework released under the MIT License, it offers organizations full control over model weights and the underlying infrastructure. This is a significant advantage for data sovereignty and regulatory compliance, allowing companies to keep AI workloads within their security boundaries, even in air-gapped environments.

The 8 NFE inference efficiency, with its focus on 'cost-effective serving,' directly translates to a more favorable TCO for proprietary infrastructures, reducing hardware requirements and long-term operational costs. For CTOs, DevOps leads, and infrastructure architects evaluating alternatives to public cloud, LongCat-Video-Avatar 1.5 represents a solution that balances high performance, flexibility, and control, aligning with deployment needs that prioritize security and cost optimization.