在 https://github.com/kijai/ComfyUI-CogVideoXWrapper 的基础上增加一个独立的模型加载器和一些其他的功能, 使用前需要先安装ComfyUI-CogVideoXWrapper
Based on https://github.com/kijai/ComfyUI-CogVideoXWrapper add model loader and some other features, you need to install ComfyUI-CogVideoXWrapper before using
添加vae encode tiling开关, 在16G显存下可以生成1024分辨率的视频不会OOM
(add vae encode tiling switch, can generate 1024 resolution video without OOM on 16G vram)
支持CogVideoX-I2V图生视频, 提供Q4量化模型
(support CogVideoX-I2V image to video, provide Q4 quantization model)
CogVideoX_5b_I2V_GGUF_Q4_0.safetensors 下载到unet文件夹 (Donwload to unet folder)
工作流在examples\workflow_I2V_GGUF_Q4_0.png (Workflow in examples\workflow_I2V_GGUF_Q4_0.png)
Prompt :The girl in the video happily puts on sunglasses.
CogVideoX5B_00064.mp4
支持CogVideoX-Fun图生视频, 提供Q4量化模型
(support CogVideoX-Fun image to video, provide Q4 quantization model)
CogVideoX_Fun_GGUF_Q4_0.safetensors 下载到unet文件夹 (Donwload to unet folder)
工作流在examples\workflow_FUN_I2V_GGUF_Q4_0.png (Workflow in examples\workflow_FUN_I2V_GGUF_Q4_0.png)
Prompt :The girl in the video happily puts on sunglasses.
1.mp4
部分层4位量化模型, 采样方式选择DPM, weight_dtype选择GGUF, cpu_offload不启用分辨率480x720在32帧下占用显存小于8G
(CogVideoX-5B 4-bit quantization model, when using it, the sampling method needs to select DPM, weight_dtype selects GGUF, and cpu_offload is not enabled)
CogVideoX_5b_GGUF_Q4_0.safetensors 下载到unet文件夹
4位量化工作流在examples\workflow_GGUF_Q4_0.png
CogVideoX-5B的模型加载器
依赖 https://github.com/kijai/ComfyUI-CogVideoXWrapper
模型下载地址
CogVideoX_5b_fp8_4em.safetensors 下载到unet文件夹
diffusion_pytorch_model.safetensors 下载到vae文件夹(记得重命名文件名方便辨认)
t5xxl_fp8_e4m3fn.safetensors 下载到clip文件夹(和sd3/flux通用,不需要重复下载)
工作流在examples/*.png
CogVideoX-5B model loader
Depends on https://github.com/kijai/ComfyUI-CogVideoXWrapper
Model download address
CogVideoX_5b_fp8_4em.safetensors Download to unet folder
diffusion_pytorch_model.safetensors Download to vae folder (remember to rename the file name for easy identification)
t5xxl_fp8_e4m3fn.safetensors Download to clip folder (universal with sd3/flux, no need to download repeatedly)
Workflow in examples/*.png