zai-org/CogVideoX-5b
69120k
1---2license: other3license_link: https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE4language:5 - en6tags:7 - cogvideox8 - video-generation9 - thudm10 - text-to-video11inference: false12---13 14# CogVideoX-5B15 16<p style="text-align: center;">17 <div align="center">18 <img src=https://github.com/THUDM/CogVideo/raw/main/resources/logo.svg width="50%"/>19 </div>20 <p align="center">21 <a href="https://huggingface.co/THUDM/CogVideoX-5b/blob/main/README_zh.md">π δΈζι
θ―»</a> | 22 <a href="https://huggingface.co/spaces/THUDM/CogVideoX-5B-Space">π€ Huggingface Space</a> |23 <a href="https://github.com/THUDM/CogVideo">π Github </a> | 24 <a href="https://arxiv.org/pdf/2408.06072">π arxiv </a>25</p>26<p align="center">27π Visit <a href="https://chatglm.cn/video?lang=en?fr=osm_cogvideo">QingYing</a> and <a href="https://open.bigmodel.cn/?utm_campaign=open&_channel_track_key=OWTVNma9">API Platform</a> to experience commercial video generation models.28</p>29 30## Demo Show31 32<!DOCTYPE html>33<html lang="en">34<head>35 <meta charset="UTF-8">36 <meta name="viewport" content="width=device-width, initial-scale=1.0">37 <title>Video Gallery with Captions</title>38 <style>39 .video-container {40 display: flex;41 flex-wrap: wrap;42 justify-content: space-around;43 }44 .video-item {45 width: 45%;46 margin-bottom: 20px;47 transition: transform 0.3s;48 }49 .video-item:hover {50 transform: scale(1.1);51 }52 .caption {53 text-align: center;54 margin-top: 10px;55 font-size: 11px;56 }57 </style>58</head>59<body>60 <div class="video-container">61 <div class="video-item">62 <video width="100%" controls>63 <source src="https://github.com/user-attachments/assets/cf5953ea-96d3-48fd-9907-c4708752c714" type="video/mp4">64 </video>65 <div class="caption">A garden comes to life as a kaleidoscope of butterflies flutters amidst the blossoms, their delicate wings casting shadows on the petals below. In the background, a grand fountain cascades water with a gentle splendor, its rhythmic sound providing a soothing backdrop. Beneath the cool shade of a mature tree, a solitary wooden chair invites solitude and reflection, its smooth surface worn by the touch of countless visitors seeking a moment of tranquility in nature's embrace.</div>66 </div>67 <div class="video-item">68 <video width="100%" controls>69 <source src="https://github.com/user-attachments/assets/fe0a78e6-b669-4800-8cf0-b5f9b5145b52" type="video/mp4">70 </video>71 <div class="caption">A small boy, head bowed and determination etched on his face, sprints through the torrential downpour as lightning crackles and thunder rumbles in the distance. The relentless rain pounds the ground, creating a chaotic dance of water droplets that mirror the dramatic sky's anger. In the far background, the silhouette of a cozy home beckons, a faint beacon of safety and warmth amidst the fierce weather. The scene is one of perseverance and the unyielding spirit of a child braving the elements.</div>72 </div>73 <div class="video-item">74 <video width="100%" controls>75 <source src="https://github.com/user-attachments/assets/c182f606-8f8c-421d-b414-8487070fcfcb" type="video/mp4">76 </video>77 <div class="caption">A suited astronaut, with the red dust of Mars clinging to their boots, reaches out to shake hands with an alien being, their skin a shimmering blue, under the pink-tinged sky of the fourth planet. In the background, a sleek silver rocket, a beacon of human ingenuity, stands tall, its engines powered down, as the two representatives of different worlds exchange a historic greeting amidst the desolate beauty of the Martian landscape.</div>78 </div>79 <div class="video-item">80 <video width="100%" controls>81 <source src="https://github.com/user-attachments/assets/7db2bbce-194d-434d-a605-350254b6c298" type="video/mp4">82 </video>83 <div class="caption">An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea.</div>84 </div>85 <div class="video-item">86 <video width="100%" controls>87 <source src="https://github.com/user-attachments/assets/62b01046-8cab-44cc-bd45-4d965bb615ec" type="video/mp4">88 </video>89 <div class="caption">In a dimly lit bar, purplish light bathes the face of a mature man, his eyes blinking thoughtfully as he ponders in close-up, the background artfully blurred to focus on his introspective expression, the ambiance of the bar a mere suggestion of shadows and soft lighting.</div>90 </div>91 <div class="video-item">92 <video width="100%" controls>93 <source src="https://github.com/user-attachments/assets/d78e552a-4b3f-4b81-ac3f-3898079554f6" type="video/mp4">94 </video>95 <div class="caption">A golden retriever, sporting sleek black sunglasses, with its lengthy fur flowing in the breeze, sprints playfully across a rooftop terrace, recently refreshed by a light rain. The scene unfolds from a distance, the dog's energetic bounds growing larger as it approaches the camera, its tail wagging with unrestrained joy, while droplets of water glisten on the concrete behind it. The overcast sky provides a dramatic backdrop, emphasizing the vibrant golden coat of the canine as it dashes towards the viewer.</div>96 </div>97 <div class="video-item">98 <video width="100%" controls>99 <source src="https://github.com/user-attachments/assets/30894f12-c741-44a2-9e6e-ddcacc231e5b" type="video/mp4">100 </video>101 <div class="caption">On a brilliant sunny day, the lakeshore is lined with an array of willow trees, their slender branches swaying gently in the soft breeze. The tranquil surface of the lake reflects the clear blue sky, while several elegant swans glide gracefully through the still water, leaving behind delicate ripples that disturb the mirror-like quality of the lake. The scene is one of serene beauty, with the willows' greenery providing a picturesque frame for the peaceful avian visitors.</div>102 </div>103 <div class="video-item">104 <video width="100%" controls>105 <source src="https://github.com/user-attachments/assets/926575ca-7150-435b-a0ff-4900a963297b" type="video/mp4">106 </video>107 <div class="caption">A Chinese mother, draped in a soft, pastel-colored robe, gently rocks back and forth in a cozy rocking chair positioned in the tranquil setting of a nursery. The dimly lit bedroom is adorned with whimsical mobiles dangling from the ceiling, casting shadows that dance on the walls. Her baby, swaddled in a delicate, patterned blanket, rests against her chest, the child's earlier cries now replaced by contented coos as the mother's soothing voice lulls the little one to sleep. The scent of lavender fills the air, adding to the serene atmosphere, while a warm, orange glow from a nearby nightlight illuminates the scene with a gentle hue, capturing a moment of tender love and comfort.</div>108 </div>109 </div>110</body>111</html>112 113## Model Introduction114 115CogVideoX is an open-source version of the video generation model originating116from [QingYing](https://chatglm.cn/video?lang=en?fr=osm_cogvideo). The table below displays the list of video generation117models we currently offer, along with their foundational information.118 119<table style="border-collapse: collapse; width: 100%;">120 <tr>121 <th style="text-align: center;">Model Name</th>122 <th style="text-align: center;">CogVideoX-2B</th>123 <th style="text-align: center;">CogVideoX-5B (This Repository)</th>124 </tr>125 <tr>126 <td style="text-align: center;">Model Description</td>127 <td style="text-align: center;">Entry-level model, balancing compatibility. Low cost for running and secondary development.</td>128 <td style="text-align: center;">Larger model with higher video generation quality and better visual effects.</td>129 </tr>130 <tr>131 <td style="text-align: center;">Inference Precision</td>132 <td style="text-align: center;"><b>FP16* (Recommended)</b>, BF16, FP32, FP8*, INT8, no support for INT4</td>133 <td style="text-align: center;"><b>BF16 (Recommended)</b>, FP16, FP32, FP8*, INT8, no support for INT4</td>134 </tr>135 <tr>136 <td style="text-align: center;">Single GPU VRAM Consumption<br></td>137 <td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> FP16: 18GB <br><b>diffusers FP16: starting from 4GB*</b><br><b>diffusers INT8(torchao): starting from 3.6GB*</b></td>138 <td style="text-align: center;"><a href="https://github.com/THUDM/SwissArmyTransformer">SAT</a> BF16: 26GB <br><b>diffusers BF16: starting from 5GB*</b><br><b>diffusers INT8(torchao): starting from 4.4GB*</b></td>139 </tr>140 <tr>141 <td style="text-align: center;">Multi-GPU Inference VRAM Consumption</td>142 <td style="text-align: center;"><b>FP16: 10GB* using diffusers</b></td>143 <td style="text-align: center;"><b>BF16: 15GB* using diffusers</b></td>144 </tr>145 <tr>146 <td style="text-align: center;">Inference Speed<br>(Step = 50, FP/BF16)</td>147 <td style="text-align: center;">Single A100: ~90 seconds<br>Single H100: ~45 seconds</td>148 <td style="text-align: center;">Single A100: ~180 seconds<br>Single H100: ~90 seconds</td>149 </tr>150 <tr>151 <td style="text-align: center;">Fine-tuning Precision</td>152 <td style="text-align: center;"><b>FP16</b></td>153 <td style="text-align: center;"><b>BF16</b></td>154 </tr>155 <tr>156 <td style="text-align: center;">Fine-tuning VRAM Consumption (per GPU)</td>157 <td style="text-align: center;">47 GB (bs=1, LORA)<br> 61 GB (bs=2, LORA)<br> 62GB (bs=1, SFT)</td>158 <td style="text-align: center;">63 GB (bs=1, LORA)<br> 80 GB (bs=2, LORA)<br> 75GB (bs=1, SFT)</td>159 </tr>160 <tr>161 <td style="text-align: center;">Prompt Language</td>162 <td colspan="2" style="text-align: center;">English*</td>163 </tr>164 <tr>165 <td style="text-align: center;">Prompt Length Limit</td>166 <td colspan="2" style="text-align: center;">226 Tokens</td>167 </tr>168 <tr>169 <td style="text-align: center;">Video Length</td>170 <td colspan="2" style="text-align: center;">6 Seconds</td>171 </tr>172 <tr>173 <td style="text-align: center;">Frame Rate</td>174 <td colspan="2" style="text-align: center;">8 Frames per Second</td>175 </tr>176 <tr>177 <td style="text-align: center;">Video Resolution</td>178 <td colspan="2" style="text-align: center;">720 x 480, no support for other resolutions (including fine-tuning)</td>179 </tr>180 <tr>181 <td style="text-align: center;">Positional Encoding</td>182 <td style="text-align: center;">3d_sincos_pos_embed</td>183 <td style="text-align: center;">3d_rope_pos_embed</td>184 </tr>185</table>186 187**Data Explanation**188 189+ When testing using the `diffusers` library, all optimizations provided by the `diffusers` library were enabled. This190 solution has not been tested for actual VRAM/memory usage on devices other than **NVIDIA A100 / H100**. Generally,191 this solution can be adapted to all devices with **NVIDIA Ampere architecture** and above. If the optimizations are192 disabled, VRAM usage will increase significantly, with peak VRAM usage being about 3 times higher than the table193 shows. However, speed will increase by 3-4 times. You can selectively disable some optimizations, including:194 195```196pipe.enable_model_cpu_offload()197pipe.enable_sequential_cpu_offload()198pipe.vae.enable_slicing()199pipe.vae.enable_tiling()200``` 201 202+ When performing multi-GPU inference, the `enable_model_cpu_offload()` optimization needs to be disabled.203+ Using INT8 models will reduce inference speed. This is to ensure that GPUs with lower VRAM can perform inference204 normally while maintaining minimal video quality loss, though inference speed will decrease significantly.205+ The 2B model is trained with `FP16` precision, and the 5B model is trained with `BF16` precision. We recommend using206 the precision the model was trained with for inference.207+ [PytorchAO](https://github.com/pytorch/ao) and [Optimum-quanto](https://github.com/huggingface/optimum-quanto/) can be208 used to quantize the text encoder, Transformer, and VAE modules to reduce CogVideoX's memory requirements. This makes209 it possible to run the model on a free T4 Colab or GPUs with smaller VRAM! It is also worth noting that TorchAO210 quantization is fully compatible with `torch.compile`, which can significantly improve inference speed. `FP8`211 precision must be used on devices with `NVIDIA H100` or above, which requires installing212 the `torch`, `torchao`, `diffusers`, and `accelerate` Python packages from source. `CUDA 12.4` is recommended.213+ The inference speed test also used the above VRAM optimization scheme. Without VRAM optimization, inference speed214 increases by about 10%. Only the `diffusers` version of the model supports quantization.215+ The model only supports English input; other languages can be translated into English during refinement by a large216 model.217 218**Note**219 220+ Using [SAT](https://github.com/THUDM/SwissArmyTransformer) for inference and fine-tuning of SAT version221 models. Feel free to visit our GitHub for more information.222 223## Quick Start π€224 225This model supports deployment using the huggingface diffusers library. You can deploy it by following these steps.226 227**We recommend that you visit our [GitHub](https://github.com/THUDM/CogVideo) and check out the relevant prompt228optimizations and conversions to get a better experience.**229 2301. Install the required dependencies231 232```shell233# diffusers>=0.30.1234# transformers>=4.44.2235# accelerate>=0.33.0 (suggest install from source)236# imageio-ffmpeg>=0.5.1237pip install --upgrade transformers accelerate diffusers imageio-ffmpeg 238```239 2402. Run the code241 242```python243import torch244from diffusers import CogVideoXPipeline245from diffusers.utils import export_to_video246 247prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."248 249pipe = CogVideoXPipeline.from_pretrained(250 "THUDM/CogVideoX-5b",251 torch_dtype=torch.bfloat16252)253 254pipe.enable_model_cpu_offload()255pipe.vae.enable_tiling()256 257video = pipe(258 prompt=prompt,259 num_videos_per_prompt=1,260 num_inference_steps=50,261 num_frames=49,262 guidance_scale=6,263 generator=torch.Generator(device="cuda").manual_seed(42),264).frames[0]265 266export_to_video(video, "output.mp4", fps=8)267```268 269## Quantized Inference270 271[PytorchAO](https://github.com/pytorch/ao) and [Optimum-quanto](https://github.com/huggingface/optimum-quanto/) can be272used to quantize the Text Encoder, Transformer and VAE modules to lower the memory requirement of CogVideoX. This makes273it possible to run the model on free-tier T4 Colab or smaller VRAM GPUs as well! It is also worth noting that TorchAO274quantization is fully compatible with `torch.compile`, which allows for much faster inference speed.275 276```diff277# To get started, PytorchAO needs to be installed from the GitHub source and PyTorch Nightly.278# Source and nightly installation is only required until next release.279 280import torch281from diffusers import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel, CogVideoXPipeline282from diffusers.utils import export_to_video283+ from transformers import T5EncoderModel284+ from torchao.quantization import quantize_, int8_weight_only, int8_dynamic_activation_int8_weight285 286+ quantization = int8_weight_only287 288+ text_encoder = T5EncoderModel.from_pretrained("THUDM/CogVideoX-5b", subfolder="text_encoder", torch_dtype=torch.bfloat16)289+ quantize_(text_encoder, quantization())290 291+ transformer = CogVideoXTransformer3DModel.from_pretrained("THUDM/CogVideoX-5b", subfolder="transformer", torch_dtype=torch.bfloat16)292+ quantize_(transformer, quantization())293 294+ vae = AutoencoderKLCogVideoX.from_pretrained("THUDM/CogVideoX-5b", subfolder="vae", torch_dtype=torch.bfloat16)295+ quantize_(vae, quantization())296 297# Create pipeline and run inference298pipe = CogVideoXPipeline.from_pretrained(299 "THUDM/CogVideoX-5b",300+ text_encoder=text_encoder,301+ transformer=transformer,302+ vae=vae,303 torch_dtype=torch.bfloat16,304)305pipe.enable_model_cpu_offload()306pipe.vae.enable_tiling()307 308prompt = "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance."309 310video = pipe(311 prompt=prompt,312 num_videos_per_prompt=1,313 num_inference_steps=50,314 num_frames=49,315 guidance_scale=6,316 generator=torch.Generator(device="cuda").manual_seed(42),317).frames[0]318 319export_to_video(video, "output.mp4", fps=8)320```321 322Additionally, the models can be serialized and stored in a quantized datatype to save disk space when using PytorchAO.323Find examples and benchmarks at these links:324 325- [torchao](https://gist.github.com/a-r-r-o-w/4d9732d17412888c885480c6521a9897)326- [quanto](https://gist.github.com/a-r-r-o-w/31be62828b00a9292821b85c1017effa)327 328## Explore the Model329 330Welcome to our [github](https://github.com/THUDM/CogVideo), where you will find:331 3321. More detailed technical details and code explanation.3332. Optimization and conversion of prompt words.3343. Reasoning and fine-tuning of SAT version models, and even pre-release.3354. Project update log dynamics, more interactive opportunities.3365. CogVideoX toolchain to help you better use the model.3376. INT8 model inference code support.338 339## Model License340 341This model is released under the [CogVideoX LICENSE](LICENSE).342 343## Citation344 345```346@article{yang2024cogvideox,347 title={CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer},348 author={Yang, Zhuoyi and Teng, Jiayan and Zheng, Wendi and Ding, Ming and Huang, Shiyu and Xu, Jiazheng and Yang, Yuanming and Hong, Wenyi and Zhang, Xiaohan and Feng, Guanyu and others},349 journal={arXiv preprint arXiv:2408.06072},350 year={2024}351}352```353 354 355