[11:09 Sat,28.September 2024 by Thomas Richter] |
Just three weeks ago, the open-source (and thus free) video AI CogVideoX was released, which already runs on consumer-grade graphics cards and can generate videos on any properly equipped home PC. A lot has happened in this short time, and the developers have added several interesting functions to CogVideoX. It now supports classic text-to-video, image-to-video, and video-to-video modes in the CogVideo Gradio app, allowing new videos to be created based on images or even other videos. Additionally, a given video can be extended using the "Extend" feature.
And the open-source nature of CogVideoX also allows further improvements in functionality and interface, which have been implemented by users. For example, Gradio now allows for "infinite" video generation, meaning a prompt is entered, and videos based on it are continuously generated—without an end. This mode is useful if you input a complex prompt but are unsatisfied with the result or the AI&s interpretation—you can simply wait in the new mode until one of the videos looks satisfactory. Another new feature is a sidebar in the interface, which displays the last 50 generated videos—practical for the "infinite" generation, to review the different versions.
Unfortunately, the generated 6-second-long videos still suffer from the same limitations as the first version of CogVideoX. They have a resolution of 720x480 pixels and a low frame rate of 8 frames per second. The free demo on CogVideoX-5B on Huggingface already integrates optional upscaling to 2880 x 1920 pixels and frame rate interpolation to 16 fps—several demo clips are available here as well. CogVideoX can also be installed on a local PC, provided a sufficiently powerful GPU with enough VRAM is available. Video-2-Video in CogVideoX CogVideoX comes in two different versions, which differ in their model size and therefore in the image quality achieved and the respective hardware requirements. By default, the smaller model CogVideoX-2B already runs on older GPUs like the GTX 1080TI with 11 GB VRAM, while the larger CogVideoX-5B model with 5 billion parameters runs on mid-range GPUs like the RTX 3060 with 12 GB VRAM—though with the help of special parameters, the memory demand of CogVideoX-5B can be further reduced, allowing it to run on graphics cards with just 5 GB of VRAM. 2 The CogVideoX-2B model was released under the generous Apache 2.0 license, while CogVideoX-5B is subject to the slightly more restrictive CogVideoX license, which grants users fee-free use of the software—for commercial use, registration and a free base license are required, allowing commercial use with up to 1 million visits per month. more infos at bei github.com deutsche Version dieser Seite: Freie Video-KI CogVideoX mit neuen Funktionen: Nonstop-Videos und Vid2Vid |