Meituan's open-weight talking-avatar model now holds a consistent face for more than 10 minutes, generates in 8 steps, and placed second only to OmniHuman-2.0 in Meituan's own human-preference vote.
- Meituan released LongCat-Video-Avatar 1.6 , an MIT-licensed audio-driven talking-avatar video model.
- Generates consistent talking videos past 10 minutes by reducing drift between chained video segments.
- A new on-policy self-distillation step trains the model on its own generated history, with a teacher that sees real video.
- DMD distillation cuts sampling from 50 to 8 steps; optional INT8 or FP8 quantization saves memory.
- Ranks second (Elo 1068.1) behind OmniHuman-2.0 in Meituan's own 33,051-vote arena, only 16 Elo above version 1.5.
- Weak spots include frequent lip-sync errors under strict review, cartoon-style inputs, and no ComfyUI node yet; code is on GitHub .
Meituan’s LongCat team has released model weights for LongCat-Video-Avatar 1.6, an open-weight system that combines an audio track with either a reference image or a text prompt to generate talking video. It can also continue an existing clip. Meituan says its post-training recipe limits error accumulation across autoregressive segments, allowing outputs beyond 10 minutes with stronger identity and motion consistency. The MIT license permits commercial use, modification, and redistribution, subject to its copyright and license-notice requirements.
Most high-performing avatar systems, including OmniHuman and HeyGen, remain closed. In Meituan’s own human-preference evaluation, version 1.6 ranks close to the strongest commercial system tested, giving developers a locally deployable alternative with inspectable weights and inference code.
Why ten-minute avatars drift #
Long avatar videos are generated as a chain of shorter segments, with each segment conditioned on the final frames of the previous one. Minor defects, such as redder lips, unstable teeth, or a wobbling mouth, become input for the next segment and can intensify over time. Meituan identifies two sources of drift: excessive generation changes within a segment and accumulated errors passed between segments.
- Calmer mouth motion: The post-training process targets strained expressions, exaggerated lip shapes, red lip coloration, and temporal instability around the mouth.
- Long-form consistency: Meituan reports clips longer than 10 minutes that preserve the subject’s appearance across chained segments.
- Faster sampling: An eight-step generation path replaces the previous 50-step process, with optional INT8 or FP8 weights and compiled kernels for lower memory use and faster inference.
The model accepts audio with a still image, audio with a text description, or an existing video that needs continuation. It supports both single-person and multi-person scenes.
Three adapters split the job #
Version 1.6 retains the LongCat-Video architecture and adds a refined post-training pipeline to Meituan’s 13.6-billion-parameter video model. The technical report combines reinforcement learning, adjusted classifier-free guidance, long-video rollouts, few-step distillation, and on-policy self-distillation. These methods are packaged as three LoRA adapters, which are compact weight sets applied over the frozen base model.
This story is for Pro members #
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.