LongAnimation: Long Animation Generation with Dynamic Global-Local Memory

https://github.com/user-attachments/assets/a3866f82-b07a-41ae-9673-2a24f7c78af4

LongAnimation: Long Animation Generation with Dynamic Global-Local Memory

Nan Chen¹, Mengqi Huang¹, Yihao Meng², Zhendong Mao^†,1
¹USTC ²HKUST ^†corresponding author

Existing studies are limited to short-term colorization by fusing overlapping features to achieve smooth transitions, which fails to maintain long-term color consistency. In this study, we propose a dynamic global-local paradigm to achieve ideal long-term color consistency by dynamically extracting global color-consistent features relevant to the current generation.

🎉🎉 Our paper, “LongAnimation: Long Animation Generation with Dynamic Global-Local Memory” accepted by ICCV 2025! Strongly recommend seeing our demo page.

Showcase

https://github.com/user-attachments/assets/8d225a9e-6e27-42bd-9638-5f4e4cb4dbf7

https://github.com/user-attachments/assets/0fee3eed-8a38-4382-bbe6-21c0cf2371e9

https://github.com/user-attachments/assets/7d87e63a-f5e6-46ba-bb1b-d457ceb0b1d8

Creative usage

Text-guided Background Generation

https://github.com/user-attachments/assets/68a5d0fb-f767-4fc8-aed6-cd798301484f

https://github.com/user-attachments/assets/7cba4d5b-b793-474d-9da4-34892853b240

https://github.com/user-attachments/assets/6787349b-6a3e-4ed1-8a6a-efc1643a4e92

A boy and a girl in different environment.

TODO List

Release the paper and demo page. Visit https://cn-makers.github.io/long_animation_web/
Release the code.

Requirements

The training is conducted on 6 A100 GPUs (80GB VRAM), the inference is tested on 1 A100 GPU.

Setup

git clone https://github.com/CN-makers/LongAnimation
cd LongAnimation

Environment

All the tests are conducted in Linux. We suggest running our code in Linux. To set up our environment in Linux, please run:

conda create -n LongAnimation python=3.10 -y
conda activate LongAnimation

bash install.sh

Checkpoints

please download the pre-trained CogVideoX-1.5 I2V checkpoints from here, and put the whole folder under pretrained_weight, it should look like ./pretrained_weights/CogVideoX1.5-5B-I2V
please download the pre-trained long video understanding model Video-XL checkpoints from here, and put the whole folder under pretrained_weight, it should look like ./pretrained_weights/videoxl
please download the checkpoint for our SketchDiT and DGLM model from here, and put the whole folder as ./pretrained_weights/longanimation.

Generate Your Animation!

To colorize the target lineart sequence with a specific character design, you can run the following command:

bash  long_animation_inference.sh

We provide some test cases in test_json folder. You can also try our model with your own data. You can change the lineart sequence and corresponding character design in the script Long_animation_inference.sh.

During the official training and testing, the --height and --weight we used were 576 and 1024 respectively. Additionally, the model can also be compatible with resolutions of 768 in length and 1360 in width respectively.

Citation:

Don't forget to cite this source if it proves useful in your research!

@misc{chen2025longanimationlonganimationgeneration,
      title={LongAnimation: Long Animation Generation with Dynamic Global-Local Memory}, 
      author={Nan Chen and Mengqi Huang and Yihao Meng and Zhendong Mao},
      year={2025},
      eprint={2507.01945},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.01945}, 
}