LongCat-Avatar: Audio-Driven Avatar Video by Meituan
LongCat-Avatar is Meituan's open audio-driven avatar model: one photo plus audio becomes lip-synced talking video. Version 1.6 holds identity past ten minutes.
LongCat-Avatar
AvatarAudio-to-VideoLip SyncVideo GenerationMeituan's audio-driven avatar video model, built on the LongCat-Video foundation. It turns one photo plus an audio track into a lip-synchronized talking video, supports Audio-Text-to-Video and Audio-Text-Image-to-Video as well as continuing an existing video, and works with a single audio stream or several. Version 1.6 keeps the face and body stable past ten minutes of generated video on a faster 8-step path.
| Developer | Meituan |
| Release Date | 2026-11 (1.6), 2026-05 (1.5), 2025-12 (original) |
| Base Model | LongCat-Video |
| License | MIT |
| Capabilities | Audio-text-to-video, audio-text-image-to-video, video continuation |
| Audio Input | Single or multiple audio streams |
| Audio Encoder | Whisper-Large (1.5) |
| Inference | 8-step distilled path: DMD LoRA early steps, OPSD LoRA later steps |
What is LongCat-Avatar?
LongCat-Avatar is Meituan's open framework for audio-driven human video, built on the LongCat-Video foundation model. Given a single photo and an audio track it produces a talking avatar with lip sync, identity preservation and full-body temporal stability over long clips, and it accepts more than one audio stream for multi-speaker input.
Version 1.5 replaced the original Wav2Vec2 encoder with Whisper-Large, which smooths out lip dynamics, and added an 8-step DMD2 distilled inference path for cheaper serving without losing visual fidelity. The release also generalizes to stylized domains, including anime, animals, multi-person interactions and object handling.
Version 1.6
1.6 targets three problems rather than new capabilities. Lip motion is cleaned up (fewer strained or over-red mouth shapes, less flicker), autoregressive error buildup is cut so clips can run past ten minutes without the face drifting, and the pipeline itself is faster. The architecture is unchanged from 1.5, so it is a checkpoint swap for existing pipelines rather than a new model family.
It still ships two task LoRAs: DMD runs the early denoising steps, OPSD the later ones, mixed with --mix_steps and --mix_alpha. The default 8-step setup uses four steps of each. An optional weight-only INT8 DiT cuts the diffusion weights from roughly 30 GB to 15 GB.
ComfyUI support has not followed yet: the community ComfyUI-LongCat-Avatar pack covers 1.5 only. See LongCat Video Avatar 1.6 for the release details and comparison videos.
Guides and workflows related to this model series.
Comments
Sign in with GitHub to join the discussion.