CogVideo and CogVideoX: open text-to-video and image-to-video models

What it supplies: This repository is for CogVideo and the CogVideoX series of video generation models. The README says CogVideoX-2B and CogVideoX-5B support text-to-video generation, and the I2V model adds image-to-video, with video continuation also listed; CogVideoX1.5-5B supports 10-second videos at higher resolution. Useful for: - Generate video from a text prompt, or from an image with the I2V model, using the diffusers cli_demo. - Run lower-memory diffusers inference with INT8 quantisation through torchao. - Fine-tune with LoRA using the provided scripts, or use CogKit and cogvideox-factory. First action: - Use Python between 3.10 and 3.12 inclusive. - Run pip install -r requirements.txt. - Follow inference/cli_demo.py for the diffusers version. - Pick a model by memory: the README lists a diffusers minimum of about 4GB for CogVideoX-2B, 5GB for CogVideoX-5B and from about 10GB for CogVideoX1.5-5B, with INT8 options lower. Prerequisites and limits: Python 3.10 to 3.12. The README's tables give per-model GPU memory, for example about 4GB minimum for CogVideoX-2B with FP16 diffusers, about 5GB minimum for CogVideoX-5B and from 10GB for CogVideoX1.5-5B, as stated there with footnoted optimisations; the SAT path needs far more. Quantised inference needs PyTorchAO or Optimum-quanto and the README says FP8 needs recent GPUs. - The README says to follow requirements.txt strictly when updating dependencies. - Memory figures depend on the diffusers or SAT route and on optimisations enabled. - Prompt quality matters: the README says a good prompt directly impacts the generated video. - Quality claims are the authors' own. The README states the repository code is released under the Apache 2.0 License and so are the CogVideoX-2B model and its modules, while the CogVideoX-5B model (including I2V and T2V) is released under the CogVideoX LICENSE on Hugging Face. Check that licence for the exact model you use, especially for commercial work. Videos you generate carry their own rights. Written from the README text only. Statements are as the README gives them; nothing was installed or run. Source inspected: https://raw.githubusercontent.com/THUDM/CogVideo/7a1af7154511e0ce4e4be8d62faa8c5e5a3532d2/README.md ; 28873 bytes ; sha256 0aaad019b64e906c54fff8359d655da4b3f6d67a0262d2d82fe34dcbe419e6f8 ; whole README text

{
  "revision": {
    "ref": "commit 7a1af7154511e0ce4e4be8d62faa8c5e5a3532d2",
    "pinned": false,
    "source_url": "https://raw.githubusercontent.com/THUDM/CogVideo/7a1af7154511e0ce4e4be8d62faa8c5e5a3532d2/README.md"
  },
  "checked_at": "2026-11-07T10:32:39.517Z",
  "usefulness": "- Generate video from a text prompt, or from an image with the I2V model, using the diffusers cli_demo.\n- Run lower-memory diffusers inference with INT8 quantisation through torchao.\n- Fine-tune with LoRA using the provided scripts, or use CogKit and cogvideox-factory.",
  "source_type": "repository",
  "availability": "available",
  "source_owner": "THUDM (Zhipu AI CogVideo team) (declared repository namespace)",
  "known_limitations": [
    "Python 3.10 to 3.12. The README's tables give per-model GPU memory, for example about 4GB minimum for CogVideoX-2B with FP16 diffusers, about 5GB minimum for CogVideoX-5B and from 10GB for CogVideoX1.5-5B, as stated there with footnoted optimisations; the SAT path needs far more. Quantised inference needs PyTorchAO or Optimum-quanto and the README says FP8 needs recent GPUs.\n\n- The README says to follow requirements.txt strictly when updating dependencies.\n- Memory figures depend on the diffusers or SAT route and on optimisations enabled.\n- Prompt quality matters: the README says a good prompt directly impacts the generated video.\n- Quality claims are the authors' own.\n\nThe README states the repository code is released under the Apache 2.0 License and so are the CogVideoX-2B model and its modules, while the CogVideoX-5B model (including I2V and T2V) is released under the CogVideoX LICENSE on Hugging Face. Check that licence for the exact model you use, especially for commercial work. Videos you generate carry their own rights.\n\nWritten from the README text only. Statements are as the README gives them; nothing was installed or run. ",
    "Rendered page text of the exact archived capture of the README page, read in full by the author; no media viewed or heard; nothing executed or installed.",
    "Source-documented steps; execution not performed. Internal affiliated review is not independent validation."
  ]
}