CogVideoX
Local AI video generator with downloadable models, NVIDIA GPU support and LoRA fine-tuning. Code and the 2B model use Apache 2.0.
CogVideoX is a family of downloadable video generation models for developers, researchers and creators who want to generate clips on their own hardware. It turns English text prompts into video, animates a supplied image and can continue an existing video. A local Gradio web interface provides a browser front end for generation.
The models work with Hugging Face Diffusers and SAT. Diffusers supports quantization to reduce GPU memory needs, with a trade-off in generation speed. Smaller models can run on NVIDIA desktop GPUs such as the GTX 1080 Ti and RTX 3060; memory requirements depend on the model and the optimizations used. The family includes models that produce clips up to 10 seconds long and image-to-video variants with flexible output resolutions.
For custom projects, CogVideoX supports LoRA fine-tuning and loading adapted models. Community integrations include a ComfyUI wrapper, while xDiT supports generation across multiple GPUs. The supplied demo also includes frame interpolation and super-resolution tools. Prompts must be in English, and the models are trained for detailed descriptions.
The code and CogVideoX-2B model are open source under Apache 2.0. The 5B model variants use the separate CogVideoX LICENSE. Hugging Face and ModelScope offer hosted demos; QingYing and an API platform provide larger commercial video models as cloud services.
