What is LongCat-Video-Avatar 1.6?
LongCat-Video-Avatar 1.6 is an open-source audio-driven avatar video model with better lip sync, long-form stability, quantization options and MIT-licensed weights.
Updated ·
LongCat-Video-Avatar 1.6: Open-Source Audio-Driven Avatar Video Model
Announcement date:
LongCat-Video-Avatar 1.6 is the latest upgrade to Meituan's open-source framework for audio-driven human video generation, building on version 1.5. It targets more natural expression, tighter lip-audio synchronization, coordinated face and body motion, and visual consistency over long generations, and it adds inference optimizations for faster deployment. In a human preference Arena with 33,051 valid votes it ranked second of six systems with a Bradley-Terry Elo score of 1068.1, behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0.
What LongCat-Video-Avatar 1.6 is
LongCat-Video-Avatar 1.6 is an open-source model for generating human avatar video from audio input. It is the latest upgrade to the LongCat-Video-Avatar framework and builds directly on version 1.5 rather than replacing its architecture.
Because it keeps the same base architecture as 1.5, the release is best understood as a quality and efficiency upgrade: the generation behavior improves, while the underlying model structure stays familiar to anyone who already worked with 1.5.
- Open-source audio-driven human video generation framework
- Latest upgrade to the line, building on version 1.5
- Same base architecture as LongCat-Video-Avatar 1.5
- Ships with two task-specific LoRA adapters: loraA for DMD and loraB for OPSD
Quality improvements over version 1.5
The 1.6 release focuses on the parts of avatar video that viewers notice first: how natural the performance looks, whether speech and mouth movement line up, and whether face and body move together convincingly. It also targets visual consistency when generation runs long.
Lip animation received specific attention. The release reduces exaggerated or strained mouth movements, overly red lip coloration, and temporal instability in the mouth region, which are common failure modes in talking-head generation.
- Improved expressive naturalness
- Improved lip-audio synchronization
- More coordinated facial and body motion
- Better visual consistency during long-form generation
- Less exaggerated or strained mouth movement and less temporal instability around the mouth
Long-form generation beyond 10 minutes
Autoregressive video continuation tends to accumulate error the longer it runs, which shows up as drift in appearance or coherence. Version 1.6 mitigates this error accumulation, enabling videos exceeding 10 minutes while preserving visual quality, identity consistency and temporal coherence.
That makes the model more plausible for longer scripted content, extended narration or other scenarios where a single continuous take matters more than a short clip.
- Mitigates error accumulation during autoregressive continuation
- Supports videos exceeding 10 minutes
- Aims to preserve visual quality, identity consistency and temporal coherence
Human preference Arena results
The release reports results from a human preference Arena with 33,051 valid individual votes. LongCat-Video-Avatar 1.6 ranked second among six evaluated systems with a Bradley-Terry Elo score of 1068.1.
It placed behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0. These are the vendor's reported Arena results; treat them as one signal among several rather than a guarantee for your own use case.
- 33,051 valid individual votes in the reported Arena
- Bradley-Terry Elo score of 1068.1
- Second of six evaluated systems
- Ranked behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0
Deployment and inference optimization
Version 1.6 introduces comprehensive inference optimizations aimed at faster and more efficient deployment. These cover both the end-to-end pipeline and model execution, so the gains apply to serving as well as to single runs.
The checkpoint package also provides base_model_int8, an optional weight-only INT8 DiT that can be enabled with --use_int8. On Hopper GPUs such as the H20, H100, H200 and H800, installing and enabling FlashAttention 3 can provide a significant inference speedup over FA2.
- End-to-end pipeline and model execution optimizations
- Optional weight-only INT8 DiT via base_model_int8 and --use_int8
- FlashAttention 3 on Hopper GPUs (H20, H100, H200, H800) for a significant speedup over FA2
- When running on more than two GPUs, set LONGCAT_A2A_WORKSPACE=1 to enable the optimized All-to-All workspace and reduce context-parallel communication overhead
Getting started and configuration notes
The model uses the same base architecture as 1.5 and includes two task-specific LoRA adapters: loraA for DMD and loraB for OPSD. The --mix_steps=N option controls how they combine: the first N denoising steps use mix_alpha * DMD + (1 - mix_alpha) * OPSD, while the remaining 8 - N steps use OPSD only.
Quantization is optional and mutually exclusive. Do not enable --use_fp8 and --use_int8 at the same time; only one quantization mode should be active. Exact hardware requirements, installation steps and default settings beyond those described in the release notes are not specified here.
- Two LoRA adapters: loraA (DMD) and loraB (OPSD)
- --mix_steps=N blends DMD and OPSD for the first N steps, then OPSD only for the remaining 8 - N steps
- Enable only one of --use_fp8 or --use_int8 at a time
- Pricing status: the model weights are released openly; no hosted pricing is stated in the source
License and usage terms
The model weights are released under the MIT License. Contributions to the repository are also licensed under the MIT License unless otherwise stated.
The license does not grant any rights to use Meituan trademarks or patents, so brand-related use needs separate consideration.
- Model weights under the MIT License
- Repository contributions under the MIT License unless otherwise stated
- No rights granted to Meituan trademarks or patents
Questions
What does LongCat-Video-Avatar 1.6 improve over version 1.5?
It improves expressive naturalness, lip-audio synchronization, coordinated facial and body motion, and visual consistency during long-form generation. Lip animation specifically reduces exaggerated or strained mouth movements, overly red lip coloration and temporal instability in the mouth region. It also mitigates error accumulation in autoregressive continuation, enabling videos exceeding 10 minutes, and adds inference optimizations for faster deployment.
How does LongCat-Video-Avatar 1.6 perform in the human preference Arena?
In a reported human preference Arena with 33,051 valid individual votes, it ranked second among six evaluated systems with a Bradley-Terry Elo score of 1068.1, behind OmniHuman-2.0 and ahead of LongCat-Video-Avatar 1.5, HeyGen, OmniHuman-1.5 and Kling Avatar 2.0.
How do I enable INT8 or FP8 quantization and FlashAttention 3?
The checkpoint package includes base_model_int8, an optional weight-only INT8 DiT enabled with --use_int8. Do not enable --use_fp8 and --use_int8 at the same time; only one quantization mode should be active. On Hopper GPUs such as the H20, H100, H200 and H800, install and enable FlashAttention 3 for a significant inference speedup over FA2.
What license applies to the LongCat-Video-Avatar 1.6 model weights?
The model weights are released under the MIT License, and repository contributions are licensed under the MIT License unless otherwise stated. The license does not grant rights to use Meituan trademarks or patents.
Original sources
Tool Info
Has this tool changed? Help us keep it up to date
LongCat-Video-Avatar 1.6 Reviews
Loading reviews…