license: apache-2.0
base_model:
- Wan-AI/Wan2.2-T2V-A14B
pipeline_tag: text-to-video
tags:
- looping
- RGB
- RGBA
- video
- generation
Loopy
Seamless Video Loop Generation via Anchored Looping Shift of Positional Embedding
Our Loopy generates high-quality looping videos with seamless transitions at loop boundaries and diverse motion. It also supports RGBA with semi-transparent effects. In the application block, all elements—including game assets and Loopy character stickers—are generated by our Loopy.
🔥 News
- [2026.8.24] Released Loopy, the Wan2.2-T2V-A14B–adapted weights and inference code are now open-sourced.
- [2026.8.24] Our technical report is available on arXiv.
📝 To-Do List
- Paper: Release the technical report on arXiv.
- Inference Code: Released the looping inference pipeline for Loopy.
- Model Weights: Release the high-noise / low-noise LoRA checkpoints.
🌟 Showcase
Seamless Looping Video Generation
| Preview Video | Preview Video |
|---|---|
For more results, please visit Our Website
🚀 Quick Start
1. Environment Setup
# Clone the project repository
git clone https://github.com/WeChatCV/Loopy.git
cd Loopy
# Create and activate Conda environment
conda create -n Loopy python=3.11 -y
conda activate Loopy
# Install dependencies
pip install -r requirements.txt
2. Model Download
Download Wan2.2-T2V-A14B
Download the LightX2V 4-step distillation LoRAs for both experts: high noise / low noise
Download Loopy (high_noise.safetensors and
low_noise.safetensors)
🧪 Usage
Write one prompt per line in a plain UTF-8 text file, then launch the batch inference:
output_path="./checkpoints"
mkdir -p $output_path
torchrun --nproc_per_node=8 generate_2.2_new.py --task t2v-A14B --size 832*480 \
--ckpt_dir Wan-AI/Wan2.2-T2V-A14B \
--dit_fsdp --t5_fsdp --ulysses_size 8 \
--frame_num 53 \
--sample_steps 4 \
--high_lightx2v_path "wan2.2_lora/wan2.2_t2v_A14b_high_noise_lora_rank64_lightx2v_4step_1217.safetensors" \
--low_lightx2v_path "wan2.2_lora/wan2.2_t2v_A14b_low_noise_lora_rank64_lightx2v_4step_1217.safetensors" \
--high_lora_path high_noise.safetensors \
--low_lora_path low_noise.safetensors \
--prompt_file prompt.txt \
--output_dir $output_path/results 2>&1 | tee -a $output_path/results.log
Or simply run the packaged script after editing the paths inside it:
bash test.sh
You can specify the weights of Wan2.2-T2V-A14B with --ckpt_dir, the LightX2V distillation
LoRAs with --high_lightx2v_path / --low_lightx2v_path, and the Loopy LoRAs with
--high_lora_path / --low_lora_path. Wan2.2-T2V-A14B is a two-expert MoE model, so the
high-noise and low-noise branches each need their own pair of LoRAs.
Generated videos are written to --output_dir, one subdirectory per prompt, each containing
fgr.mp4.
Prompt Writing Tip: Describe a subject whose motion is naturally periodic or continuous — falling snow, drifting clouds, flowing water, a walking animal, a rotating object. State the visual style and the shot type (close-up, medium shot, wide shot) as well. Prompts support both Chinese and English input.
# An example of prompt.
An Arctic fox leaps nimbly through the snow while hunting, its white fur blending seamlessly with the snowflakes. Realistic style; a scene of winter wildlife.
🤝 Acknowledgements
This project is built upon the following excellent open-source projects:
- Wan2.2 (base video generation model)
- LightX2V (inference acceleration)
- DiffSynth-Studio (training/inference framework)
We sincerely thank the authors and contributors of these projects.
✏️ Citation
If you find our work helpful for your research, please consider citing our paper:
@article{haotiandong2026loopy,
title = {Loopy: Seamless Video Loop Generation via Anchored Looping Shift of Positional Embedding},
author = {Haotian Dong, Wenjing Wang, Chen Li, Jing Lyu, Xin Wang, Di Lin},
journal = {arXiv preprint},
year = {2026}
}
📬 Contact Us
If you have any questions or suggestions, feel free to reach out via GitHub Issues. We look forward to your feedback!