Video Generation of Large Models

MAGI-1 is the world's first large autoregressive video generation model open-sourced by Sand AI. It adopts autoregressive architecture, generates smooth and natural videos by predicting video sequences block by block, and supports unlimited scaling and one-shot-to-the-end long video generation. The native resolution of the model can be up to 1440x2568, and the generated video has smooth movements and realistic details, and has the ability of controlled generation, which can realize smooth scene transitions and fine-grained control through chunking cues. MAGI-1 Features High-performance video generation: MAGI-1 can generate high-quality video clips in a short period of time, such as generating a 5-second video in just 3 seconds, generating...

Aliyun announced on February 25th that its visual generation base model Wan Phase 2.1 (Wan) is open source. The open source adopts the loosest Apache2.0 protocol, and all the inference codes and weights of 14B and 1.3B parameter specifications are open source, while supporting text-generated video and graph-generated video tasks, which can be downloaded and experienced by global developers in Github, HuggingFace and Magic Ride communities. According to the introduction, the 14B Universal Phase model has outstanding performance in instruction following, complex motion generation, physical modeling, text-video generation, etc., and has been evaluated in the review set VBe...

Beanbag video generation model release site. Jiang Jialing photo People's Daily Shenzhen, September 24 (Jiang Jialing) September 24, byte jumping under the volcano engine in Shenzhen held AI innovation tour, the scene released beanbag video generation - PixelDance, beanbag video generation - Seaweed two big models for the enterprise market to open the invitation to test. According to reports, most of the previous video generation models can only complete simple instructions. The new beanbag video generation model allows the video to switch freely between large dynamics and transport mirrors, with zoom, surround, target following and other multi-camera language capabilities, and with professional-grade light and shadow layout and color...

Ali Tongyi will release a big model of video generation during the 2024 Yunqi Conference (September 19th to September 21st). At present, Tongyi App channel and PC terminal of Tongyi Wanxiang have on-line "Video Generation" portal, which only supports users to make reservation, but is not open for use. The content of the page shows that the newly released video generation model is independently researched and developed by Tongyi Labs, and contains two creation modes: "text-generated video" and "map-generated video". After successful reservation, Ali Tongyi will notify users when they are qualified to use it, and give them 66 inspiration values (valid for 30 days). ...

At the 2024 World Artificial Intelligence Conference (WAIC), SenseTime launched Vimi, a large model for generating controllable human videos. The Vimi model uses SenseTime's advanced large model technology to generate human videos consistent with the target action from a single photo, achieving precise expression and body control. The model supports multiple driving methods, including video, animation, sound and text. It has years of accumulated facial tracking technology and precise control of details, and can generate highly consistent video content with harmonious light and shadow. Vimi's stability is particularly outstanding, and it can generate single-lens videos of more than 1 minute...

SenseTime released the first "controllable" character video generation model Vimi at the World Artificial Intelligence Conference (WAIC). It can generate character videos consistent with the target action through a photo of any style, and supports multiple driving methods. It can be driven by existing character videos, animations, sounds, text and other elements. Unlike image expression control technologies that can only control head expressions and movements, SenseTime claims that Vimi can not only achieve precise character expression control, but also control the natural body changes of the characters in the photo in the half-body area, and automatically generate hair and clothing that match the characters...

Kuaishou's "Keling" video generation model is officially launched. Officials said that the Keling model was developed by the Kuaishou AI team. It uses a similar technical route to Sora, and claims to combine multiple self-developed technical innovations and benchmark Sora's effects. According to reports, the Keling model has the following advantages: It can generate large-scale reasonable movements and simulate the characteristics of the physical world. It has a strong ability to combine concepts and imagination. The generated video resolution is 1080p, the duration is 2 minutes (frame rate 30fps), and it supports free adjustment of aspect ratio. According to Kuaishou, the Keling model has been used in the Kuaiying App...