CogVideo/CogVideoX is an open-source project for text-to-video and image-to-video generation, with multiple model variants (2B, 5B, and 1.5 versions) and SAT/diffusers-based workflows. It provides release notes and a detailed quick-start, model table, and download links for Diffusers hosting. Latest release is CogVideoX-1.0 (2024-11-08).
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
CogVideo & CogVideoX is an open-source project for text-to-video and image-to-video generation. It offers multiple model variants, including CogVideoX-2B, CogVideoX-5B, CogVideoX-5B-I2V, and CogVideoX1.5-5B, with releases such as CogVideoX-1.0. The repository provides guidance for SAT weights, Diffusers-based inference, and finetuning workflows. The project updates include new tools like CogKit for fine-tuning and inference across CogVideoX series.
How it works
The README details model variants with different resolutions, frame counts, and precision options. It describes SAT-based and Diffusers-based inference paths, and mentions support for quantized inference via diffusers-torchao. It includes performance estimates for single- and multi-GPU configurations and compares BF16/FP16/FP32/INT8/INT16-like modes across models. It also notes a multi-task capability progression: text-to-video generation, video continuation, and image-to-video generation (for CogVideoX-5B-I2V).
Getting started
Quick start sections cover prompts optimization, SAT usage, and Diffusers inference. The Diffusers installation snippet is:
pip install -r requirements.txt
The SAT and inference workflows are documented under respective sections such as sat_demo and inference/cli_demo.py. Install and usage commands are present in the README as shown above.
Recent releases
Latest release: v1.0 CogVideoX-1.0 (2024-11-08). The notes specify that if you are using CogVideoX-2B/5B/5B-I2V models, you should use the SAT code release CogVideoX-1.0, and for CogVideoX1.5 you should use CogVideoX1.5 release code. The README also lists model entries and release dates for CogVideoX1.5-5B (Nov 8, 2024) and CogVideoX-2B (Aug 6, 2024), CogVideoX-5B (Aug 27, 2024), and CogVideoX-5B-I2V (Sept 19, 2024).
Traction
Stars: 12931. (There is no 7d or 1d stars data provided in FACTS, so this section is omitted per rules.)
Behind the repo
The repository mentions integration points with HuggingFace Diffusers releases and external platforms, including links to HuggingFace spaces and ModelScope for model download and demo experiences. It also references CogKit as a forthcoming toolkit for fine-tuning and inference across CogVideoX and CogVideo series.
Caveats
License: Apache-2.0. Language: Python. Created: 2022-05-29. Last push: 2025-11-04. The project details multiple model variants with various memory, speed, and precision considerations, and notes explicit supported/inferred configurations (e.g., video resolutions, frame counts, and token limits) in the model introduction table.