CogVideoX
Open text-to-video and image-to-video models from Zhipu AI / Tsinghua.
- Stars
- 13.1K
- Forks
- 1.4K
- Open issues
- 115
- Last commit
- 2025-11-04
Live from the GitHub API · updated 2026-10-15
What is CogVideoX?
CogVideoX is a family of open video models (2B and 5B, plus image-to-video variants) that run in Diffusers. With CPU offloading and quantization the smaller models fit consumer GPUs.
Key features
- Text-to-video, image-to-video and video-to-video
- Diffusers pipelines with memory optimizations
- LoRA fine-tuning scripts
System requirements
- Platforms
- Linux · Windows
- GPU
- NVIDIA
- VRAM
- ~12–24 GB depending on model and optimizations (see the README table)
How to install CogVideoX
Summarized from the project’s README. Projects change quickly — check the GitHub page for the latest instructions.
Diffusers
- 1Install Diffusers.
pip install --upgrade diffusers transformers accelerate imageio-ffmpeg - 2Generate a clip in Python.
import torch from diffusers import CogVideoXPipeline from diffusers.utils import export_to_video pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16) pipe.enable_model_cpu_offload() video = pipe(prompt="A panda playing guitar in a bamboo forest", num_frames=49).frames[0] export_to_video(video, "output.mp4", fps=8)
Strengths
- Runs on consumer GPUs with offloading
- Good Diffusers support
Limitations
- Short, low-resolution clips
- Weights for some sizes use a custom license
CogVideoX alternatives compared
| Software | Stars | License | Platforms | VRAM |
|---|---|---|---|---|
| CogVideoX Open text-to-video and image-to-video models from Zhipu AI / Tsinghua. | 13.1K | Apache-2.0 | ~12–24 GB depending on model and optimizations (see the README table) | |
| Wan 2.2 Alibaba’s open-weight video generation models (text, image and speech to video). | 17.9K | Apache-2.0 | 24 GB for TI2V-5B at 720p; 80 GB-class GPUs (or offloading) for A14B models | |
| LTX-Video Fast, real-time-capable open video generation from Lightricks. | 11.1K | Apache-2.0 | 12 GB+ for smaller variants; more for 13B models | |
| HunyuanVideo Tencent’s 13B-parameter open video foundation model. | 12.6K | Tencent Hunyuan Community License | 45 GB minimum, 60–80 GB recommended for 720p (per README); less with community quantizations | |
| FramePack Long image-to-video generation on laptop GPUs with as little as 6 GB VRAM. | 17.3K | Apache-2.0 | 6 GB minimum | |
| AnimateDiff Animate your Stable Diffusion checkpoints with plug-in motion modules. | 12.3K | Apache-2.0 | 8–12 GB |
Related models
FAQ
Is CogVideoX free?
Yes — CogVideoX is open-source software released under the Apache-2.0 license. Check the license (and any model weight licenses) before commercial use.
Does CogVideoX run on Windows, Mac and Linux?
CogVideoX supports: Linux, Windows.
What hardware do I need for CogVideoX?
NVIDIA. VRAM: ~12–24 GB depending on model and optimizations (see the README table)
How do I install CogVideoX?
1. Install Diffusers. 2. Generate a clip in Python.
What are the best CogVideoX alternatives?
Popular alternatives include Wan 2.2, LTX-Video, HunyuanVideo, FramePack. See the comparison table on this page.
Can I use CogVideoX online without installing it?
If you don’t have a large GPU, AiLoft runs Veo 3.1 and Kling 2.5 in the cloud — generate and compare clips in the browser.
No 24 GB GPU? Generate AI video online
Run Veo 3.1 and Kling 2.5 in the cloud and compare clips side by side — pay per clip.