Meituan open-sources LongCat-Video-Avatar 1.6 with 10-min stable video
linoy_tsaban · x · 2026-11-18
Meituan's LongCat team released LongCat-Video-Avatar 1.6 on Hugging Face under MIT license, an open-source audio-driven human video generation model.
Key upgrades over 1.5:
- More natural, stable lip animation (less exaggerated mouth movement, oversaturated lips, temporal instability)
- Stable long-form generation: mitigates autoregressive error accumulation, enabling videos over 10 minutes with consistent identity and temporal coherence
- Faster end-to-end inference via pipeline and execution optimizations
Human evaluation comparisons against 1.5 and four commercial systems are included, with a technical report, weights, and an online Space all available. Blogger linoytsaban demoed it using portraits from a Krea 2 LoRA with Drama 3 audio.
Related event: Meituan Open-Sources LongCat-Video-Avatar 1.6 Audio-Driven Avatar Model(2 posts)→
More from Multimodal
- Selective-color portrait with Gemini 4 Argon: peeled layer revealing vivid eyes — AnswerVinod · 2026-11-18
- HeyGen launches Avatar 6: 84% of people can't tell its avatars from real humans — HeyGen · 2026-11-18
- Midjourney's Edit Model Now Supports Moodboards, Style Codes and Style References — midjourney · 2026-11-18
- MOSS-Image FP16-fixed weights bring speed boost to aging Turing GPUs — mkrGG · 2026-11-18
- One JSON prompt, three tools: Claude vs ChatGPT vs Figma Make for pixel-perfect UI — Tegadesigns · 2026-11-18
- Kling Motion Control Keeps Importing Reference Image Background Into Generated Videos — Flat-Scholar1878 · 2026-11-18