Skip to content

[SIGGRAPH ASIA 2024 TCS] AnimateLCM: Computation-Efficient Personalized Style Video Generation without Personalized Video Data

License

Notifications You must be signed in to change notification settings

G-U-N/AnimateLCM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

83 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

⚡️AnimateLCM: Computation-Efficient Personalized Style Video Generation without Personalized Video Data

[Paper] [Project Page ✨] [Demo in 🤗Hugging Face] [Pre-trained Models] [Civitai] visitors

by Fu-Yun Wang, Zhaoyang Huang📮, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, Hongsheng Li📮

Example 1 Example 2 Example 3
GIF 1 GIF 2 GIF 3

If you use any components of our work, please cite it.

@article{wang2024animatelcm,
  title={AnimateLCM: Accelerating the Animation of Personalized Diffusion Models and Adapters with Decoupled Consistency Learning},
  author={Wang, Fu-Yun and Huang, Zhaoyang and Shi, Xiaoyu and Bian, Weikang and Song, Guanglu and Liu, Yu and Li, Hongsheng},
  journal={arXiv preprint arXiv:2402.00769},
  year={2024}
}

News

  • [2024.05]: 🔥🔥🔥 We release the training script for accelerating Stable Video Diffusion.
  • [2024.03]: 😆😆😆 We release the AnimateLCM-I2V and AnimateLCM-SVD for fast image animation.
  • [2024.02]: 🤗🤗🤗 Release pretrained model weights and Huggingface Demo.
  • [2024.02]: 💡💡💡 Technical report is available on arXiv.

Here is a screen recording of usage. Prompt:"river reflecting mountain"

case1x2

Introduction

Animate-LCM is a pioneer work and exploratory on fast animation generation following the consistency models, being able to generate animations in good quality with 4 inference steps.

It relies on the decoupled learning paradigm, firstly learning image generation prior and then learning the temporal generation prior for fast sampling, greatly boosting the training efficiency.

The High-level workflow of AnimateLCM can be

comparison

Demos

We have launched lots of demo videos generated by Animate-LCM on the Project Page. Generally speaking, AnimateLCM works for fast, text-to-video, control-to-video, image-to-video, video-to-video stylization, and longer video generation.

comparison

Models

So far, we have released three models for usage

  • Animate-LCM-T2V: A spatial LoRA weight and a motion module for personalized video generation. Some trying from the community point out that the motion module is also compatible with many personalized models tuned for LCM, for example Dreamshaper-LCM.

  • AnimateLCM-SVD-xt. I provide AnimateLCM-SVD-xt and AnimateLCM-SVD-xt 1.1, which are tuned from SVD-xt and SVD-xt 1.1 respectively. They work for high-resolution image animation with 25 frames with 1~8 steps. You can try it with the Hugging Face Demo. Thanks to the Hugging Face team for providing the GPU grants.

  • AnimateLCM-I2V. A spatial LoRA weight and a motion module with an additional image encoder for personalized image animation. It is our trying to directly train an image animation model for fast sampling without any teacher models. It can generate animations with a personalized image with 2~4 steps. Yet due to the training resources is very limited, it is not as stable as I would like (Just like most I2V models built on Stable-Diffusion-v1-5, they generally not very stable for generation).

Install & Usage Instruction

We split the animatelcm_sd15 and animatelcm_svd into two folders. They are based on different environments. Please refer to README_animatelcm_sd15 and README_animatelcm_svd for instructions.

Usage Tips

  • AnimateLCM-T2V:

    • 4 steps can generally work well. For better quality, apply 6~8 inference steps to improve the generation quality.
    • CFG scale should be set between 1~2. Set CFG=1 can reduce the sampling cost by half. However, generally, I would prefer using CFG 1.5 and setting proper negative prompts for sampling to achieve better quality.
    • Set the video length to 16 frames for sampling. This is the length that the model trained with.
    • The models should work with IP-Adapter, ControlNet, and lots of adapters tuned for Stable Diffusion in a zero-shot manner. If you hope for better results of combination, you can try to tune them together by applying the teacher-free adaptation script I provide. It will not corrupt the sampling speed.
  • AnimateLCM-I2V:

    • 2-4 steps should work for personalized image animation.

    • In most cases, the model does not need CFG values. Just set the CFG=1 to reduce inference cost.

    • I additionally set a motion scale hyper-parameter. Set it to 0.8 as the default choice. If you set it to 0.0, you should always obtain static animations. You can increase the motion scale for larger motions, but that will sometimes cause generation failure.

    • The typical workflow can be:

      • Using your personalized image models to generate an image with good quality.
      • Applying the generated image as input and reusing the same prompt for image animation.
      • You can even further apply AnimateLCM-T2V to refine the final motion quality.
  • AnimateLCM-SVD:

    • 1-4 steps should work.
    • SVD requires two CFG values. CFG_min and CFG_max. By default, CFG_min is set to 1. Slightly adjusting CFG_max between [1, 1.5] will obtain good results. Again, just setting it to 1 to reduce the inference cost.
    • For other hyper-parameters of AnimateLCM-SVD-xt, please just follow the original SVD design.

Related Notes

Comparison

Screen recording of AnimateLCM-T2V. Prompt: "dog with sunglasses".

case2x2

comparison

Contact & Collaboration

I am open to collaboration, but not to a full-time intern. If you find some of my work interesting and hope for collaboration/discussion in any format, please do not hesitate to contact me.

📧 Email: fywang@link.cuhk.edu.hk

Acknowledge

I would thank AK for broadcasting our work and the hugging face team for providing help in building the gradio demo and storing the models. Would thank the Dhruv Nair for providing help in diffusers.

About

[SIGGRAPH ASIA 2024 TCS] AnimateLCM: Computation-Efficient Personalized Style Video Generation without Personalized Video Data

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages