CogVideoX 2B and 5B: local setup with Pinokio

Learn the CogVideoX setup through Pinokio, model selection and video generation. This September 2024 tutorial requires an Nvidia GPU.

Player not loading? Watch on YouTube

Uploaded on September 4, 2024, this tutorial covers CogVideoX-2B and 5B through Pinokio. It describes the setup available then; no Pinokio version number or separate recording date is supplied. The speaker presents CogVideoX as a local AI tool for generating five-second clips from text prompts.

The walkthrough installs Pinokio, finds CogVideoX in its discovery page, downloads the project and installs dependencies. Pinokio download links include Windows, Mac and Linux, but the demonstrated method requires an Nvidia GPU. The speaker reports installation errors and suggests restarting the application or stopping and starting the server to open the Gradio interface.

For full precision, the speaker cites about 18 GB of VRAM for 2B and 26 GB for 5B. A claim that the workflow can use as little as 6 GB does not come with a specific configuration in the transcript. The licensing discussion identifies Apache 2.0 for 2B and a more restrictive CogVideoX license for 5B.

The interface lets users enter a prompt, choose 2B or 5B and select BF16 or float16. Optional prompt enhancement uses an OpenAI API key, so that step involves an external service. The speaker warns that the first generation downloads a model that can exceed 12 GB and estimates generation at 15 minutes or more, depending on the PC. The tutorial also introduces a video-to-video tab.