88Z AI, 22 May — Meituan's LongCat team releases LongCat-Video-Avatar 1.5, significantly improving lip sync and video stability by upgrading the audio encoder, optimizing the data system, and introducing GRPO alignment. The model uses DMD distillation for 8-step generation, boosting efficiency by approximately 15 times, and advancing digital human technology from open-source SOTA to commercial-grade applications.
At a glance
- Introduces frame-level GRPO alignment and Whisper-large encoder to comprehensively improve lip sync and video stability.
- Uses DMD distillation for 8-step generation, boosting efficiency by approximately 15 times to accelerate commercial adoption.
The update
Meituan's LongCat team has officially open-sourced the LongCat-Video-Avatar 1.5 digital human video model, marking a transition from open-source state-of-the-art to commercial-grade applications. This upgrade integrates the Whisper-large audio encoder and establishes a high-quality multi-scenario data system. By introducing frame-level GRPO preference alignment, the model achieves comprehensive improvements in lip-sync accuracy, physical plausibility, long-video stability, and multi-person interaction. Furthermore, the model employs DMD distillation to compress the generation process to just 8 steps, increasing efficiency by approximately 15 times. This significantly reduces computational costs and improves response speed, providing strong technical support for the large-scale deployment of digital humans.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 龙猫LongCatSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.