LongCat-Video creates multi-minute videos and talking avatars with open AI
Meituan project using open AI to create videos several minutes long and talking avatars under the MIT license.
LongCat-Video is an open artificial intelligence model created by the Chinese company Meituan. It turns a text description, an image, or the beginning of a video into new scenes. It can also continue a video and create videos several minutes long, at 720p and 30 frames per second. The LongCat-Video-Avatar 1.5 version uses a photo and audio to make a person speak, with synchronized mouth movements. The code and models are under the MIT license.
For video creators, this makes it possible to create scenes and extend a video using the same model. Avatar 1.5 also offers a way to animate people from a photo and audio. Because the model is open, it can run on your own server, but that requires powerful equipment: the guide highlights the use of NVIDIA graphics cards and plenty of disk space. The project has not yet been tested on the server mentioned in the guide.
To get started, you need Linux with NVIDIA and CUDA, and must set up the programming environment specified by the project. Then download only the model files needed for the task: the base model to generate or continue videos, or Avatar 1.5 to create talking avatars. There are sample commands and an interface that runs in a browser. Before planning to use it, check the disk space required and note that the README does not specify the minimum graphics card memory.
The link opens the original content; some require a subscription.
Go further on INEMA.CLUB
More news
- 15/10VoiceStudio: a free, open source local alternative to ElevenLabsThe free, open source program runs ElevenLabs features locally, including voice cloning, voice from a description, dubbing, and transcription; it proposes testing TTS engines in Portuguese with Nei's voice.
- 15/10VoiceStudio: voice cloning, dubbing, and transcription on your own computerVoiceStudio provides local voice cloning, dubbing, and transcription as an alternative to ElevenLabs.
- 15/10sala-multiplayer: people, Claude, Codex, and Telegram in one conversationThe project brings people, Claude, Codex, and Telegram together in a conversation with a live dashboard.
- 15/10robotica-so101 simulates the SO-101 arm before a real-table testThe project simulates the SO-101 arm picking up and releasing objects before testing the movement on a real table.
- 15/10reel3d-teste creates animated 3D videos with code and realistic lightingThe project creates 3D scenes of a bedroom and a house with realistic lighting, without using video AI.