DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors

     
           

Jinbo Xing, Menghan Xia*, Yong Zhang, Haoxin Chen, Wangbo Yu,
Hanyuan Liu, Xintao Wang, Tien-Tsin Wong*, Ying Shan

(* corresponding authors)

From CUHK and Tencent AI Lab.

🔆 Introduction

🔥🔥 Training / Fine-tuning code is available NOW!!!

🔥 We 1024x576 version ranks 1st on the I2V benchmark list from VBench!
🔥 Generative frame interpolation / looping video generation model weights (320x512) have been released!
🔥 New Update Rolls Out for DynamiCrafter! Better Dynamic, Higher Resolution, and Stronger Coherence!
🤗 DynamiCrafter can animate open-domain still images based on text prompt by leveraging the pre-trained video diffusion priors. Please check our project page and paper for more information.

👀 Seeking comparisons with Stable Video Diffusion and PikaLabs? Click the image below.

1.1. Showcases (576x1024)

1.2. Showcases (320x512)

1.3. Showcases (256x256)

"bear playing guitar happily, snowing"		"boy walking on the street"

2. Applications

2.1 Storytelling video generation (see project page for more details)

2.2 Generative frame interpolation

Input starting frame	Input ending frame	Generated video

2.3 Looping video generation

📝 Changelog

• [2024.05.24]: 🔥🔥 Release WebVid10M-motion annotations.
• [2024.05.05]: Release training code.
• [2024.03.14]: Release generative frame interpolation and looping video models (320x512).
• [2024.02.05]: Release high-resolution models (320x512 & 576x1024).
• [2023.12.02]: Launch the local Gradio demo.
• [2023.11.29]: Release the main model at a resolution of 256x256.
• [2023.11.27]: Launch the project page and update the arXiv preprint.

🧰 Models

Model	Resolution	GPU Mem. & Inference Time (A100, ddim 50steps)	Checkpoint
DynamiCrafter1024	576x1024	18.3GB & 75s (perframe_ae=True)	Hugging Face
DynamiCrafter512	320x512	12.8GB & 20s (perframe_ae=True)	Hugging Face
DynamiCrafter256	256x256	11.9GB & 10s (perframe_ae=False)	Hugging Face
DynamiCrafter512_interp	320x512	12.8GB & 20s (perframe_ae=True)	Hugging Face

Currently, our DynamiCrafter can support generating videos of up to 16 frames with a resolution of 576x1024. The inference time can be reduced by using fewer DDIM steps.

GPU memory consumed on RTX 4090 reported by @noguchis in Twitter: 18.3GB (576x1024), 12.8GB (320x512), 11.9GB (256x256).

⚙️ Setup

Install Environment via Anaconda (Recommended)

conda create -n dynamicrafter python=3.8.5
conda activate dynamicrafter
pip install -r requirements.txt

💫 Inference

1. Command line

Image-to-Video Generation

1. Download pretrained models via Hugging Face, and put the model.ckpt with the required resolution in checkpoints/dynamicrafter_[1024|512|256]_v1/model.ckpt.
2. Run the commands based on your devices and needs in terminal.

  # Run on a single GPU:
  # Select the model based on required resolutions: i.e., 1024|512|320:
  sh scripts/run.sh 1024
  # Run on multiple GPUs for parallel inference:
  sh scripts/run_mp.sh 1024

Generative Frame Interpolation / Looping Video Generation

Download pretrained model DynamiCrafter512_interp and put the model.ckpt in checkpoints/dynamicrafter_512_interp_v1/model.ckpt.

  sh scripts/run_application.sh interp # Generate frame interpolation
  sh scripts/run_application.sh loop   # Looping video generation

2. Local Gradio demo

Image-to-Video Generation

1. Download the pretrained models and put them in the corresponding directory according to the previous guidelines.
2. Input the following commands in terminal (choose a model based on the required resolution: 1024, 512 or 256).

  python gradio_app.py --res 1024

Generative Frame Interpolation / Looping Video Generation

Download the pretrained model and put it in the corresponding directory according to the previous guidelines.

  python gradio_app_interp_and_loop.py 

💥 Training / Fine-tuning

Image-to-Video Generation

0. Download the WebVid Dataset, and important items in .csv are page_dir, videoid, and name.
1. Download the pretrained models and put them in the corresponding directory according to the previous guidelines.
2. Change <YOUR_SAVE_ROOT_DIR> path in training_[1024|512]_v1.0/run.sh
3. Carefully check all paths in training_[1024|512]_v1.0/config.yaml, including model:pretrained_checkpoint, data:data_dir, and data:meta_path.
4. Input the following commands in terminal (choose a model based on the required resolution: 1024 or 512).

We adopt DDPShardedStrategy by default for training, please make sure it is available in your pytorch_lightning.

  sh configs/training_1024_v1.0/run.sh ## fine-tune DynamiCrafter1024

5. All the checkpoints/tensorboard record/loginfo will be saved in <YOUR_SAVE_ROOT_DIR>.

🎁 WebVid-10M-motion annotations (~2.6M)

The annoations of our WebVid-10M-motion is available on Huggingface Dataset. In addition to the original annotations, we add three more motion-related annotations: dynamic_confidence, dynamic_wording, and dynamic_source_category. Please refer to our supplementary document (Section D) for more details.

🤝 Community Support

1. ComfyUI and pruned models (bf16): ComfyUI-DynamiCrafterWrapper (Thanks to kijai)

Model	Resolution	GPU Mem.	Checkpoint
DynamiCrafter1024	576x1024	10GB	Hugging Face
DynamiCrafter512_interp	320x512	8GB	Hugging Face

2. ComfyUI: ComfyUI-DynamiCrafter (Thanks to chaojie)

3. ComfyUI: ComfyUI_Native_DynamiCrafter (Thanks to ExponentialML)

4. Docker: DynamiCrafter_docker (Thanks to maximofn)

👨‍👩‍👧‍👦 Crafter Family

VideoCrafter1: Framework for high-quality video generation.

ScaleCrafter: Tuning-free method for high-resolution image/video generation.

TaleCrafter: An interactive story visualization tool that supports multiple characters.

LongerCrafter: Tuning-free method for longer high-quality video generation.

MakeYourVideo, might be a Crafter:): Video generation/editing with textual and structural guidance.

StyleCrafter: Stylized-image-guided text-to-image and text-to-video generation.

😉 Citation

Please consider citing our paper if our code and dataset annotations are useful:

@article{xing2023dynamicrafter,
  title={DynamiCrafter: Animating Open-domain Images with Video Diffusion Priors},
  author={Xing, Jinbo and Xia, Menghan and Zhang, Yong and Chen, Haoxin and Yu, Wangbo and Liu, Hanyuan and Wang, Xintao and Wong, Tien-Tsin and Shan, Ying},
  journal={arXiv preprint arXiv:2310.12190},
  year={2023}
}

🙏 Acknowledgements

We would like to thank AK(@_akhaliq) for the help of setting up hugging face online demo, and camenduru for providing the replicate & colab online demo, and Xinliang for his support and contribution to the open source project.

📢 Disclaimer

This project strives to impact the domain of AI-driven video generation positively. Users are granted the freedom to create videos using this tool, but they are expected to comply with local laws and utilize it responsibly. The developers do not assume any responsibility for potential misuse by users.

---