Radar Telegram alerts Open Radar

Meituan has released LongCat-Video-Avatar 1.6 - a new version of talking, god, singing avatars that last more than 10 minutes.

תרגם לעברית

Метаверсище и ИИще · · Source

🇨🇳 China · Map

Meituan has released LongCat-Video-Avatar 1.6 - a new version of talking, god, singing avatars that last more than 10 minutes.

Meituan has released LongCat-Video-Avatar 1.6 - a new version of talking, god, singing avatars that last more than 10 minutes.

Open source crap and sticks are usually very boring to comment on. But there is reason to relax a little. Authors - Meituan Company - China's largest "local life" platform (local services super-app): food delivery, "instant" retail, coupons/orders for offline services, hotels and travel, transportation/bike sharing and business services. The company generates revenue from booking commissions, advertising within the ecosystem and paid B2B tools for sellers.

That's right, food delivery people are now training basic models to create video.

And these messengers released, under the MIT license, the weights of the new LongCat-Video-Avatar 1.6 model, which turns a photo and audio track into a talking or singing character. The upgrade compared to version 1.5 focuses on natural lip sync, maintaining identity in videos longer than 10 minutes, and speeding up inferences.

The model is built on the basis of the basic LongCat-Video messenger and supports the creation of one character and several characters (audio-text-to-video, audio-image-text-to-video and audio-driven continuation modes). In version 1.6, the developers suppressed typical artifacts of the previous version - hypertrophic expression, unnaturally red lips and accumulation of errors during autoregressive video continuation. Due to pipeline optimization, the inference corresponds to 8 denoising steps with mixing of two LoRA adapters: in the first N steps (--mix_steps=N), DMD LoRA (loraA) and OPSD LoRA (loraB) are mixed with the mix_alpha coefficient, and the remaining 8 - N steps are perfected by OPSD.

To save memory, the kit includes a DiT checkpoint base_model_int8 (in weight only INT8, enabled by the --use_int8 flag), and FlashAttention-2 is enabled by default (FlashAttention-3 and xformers are also supported). Arena blind human comparison (33,051 votes between 6 systems) LongCat-Video-Avatar 1.

More on Radar

Want to see every report in real time?

Hundreds of sources worldwide, translated into your language, with alerts for what matters to you.

Sign up free
Share on WhatsApp Get alerts on Telegram