Wan2.2-I2V-A14B-Moe-Distill-Lightx2v

Overview

Wan2.2-I2V-A14B-Moe-Distill-Lightx2v is an advanced image-to-video generation model built upon the Wan2.2-I2V-A14B foundation. This approach allows the model to generate videos with significantly fewer inference steps (4 steps, 2 steps for high noise and 2 steps for low noise) and without classifier-free guidance, substantially reducing video generation time while maintaining high quality outputs.

This version has the following features:

1. We found that the training challenge of Wan2.2 lies in high noise; therefore, we have focused on the two-step training for the high-noise model. Compared with the previous version, we have adopted several new strategies, which have improved the consistency and dynamics of the model.
2. We found that the low-noise model can achieve good results by directly using the LoRA from Wan2.1; thus, the low-noise model still adopts the old Wan2.1 LoRA.

Inference

Our inference framework utilizes lightx2v, a highly efficient inference engine that supports multiple models. This framework significantly accelerates the video generation process while maintaining high quality output.

Edit the I2V launch script: set lightx2v_path, model_path, CUDA_VISIBLE_DEVICES, and the input/output paths.

For the full distilled checkpoints in this repository, use configs/distill/wan22/wan_moe_i2v_distill_model.json. Set high_noise_original_ckpt and low_noise_original_ckpt to the local files downloaded from distill_models/high_noise_model/ and distill_models/low_noise_model/, respectively. Keep model_path pointed at the complete Wan2.2-I2V-A14B base model directory.

Run from the LightX2V repository root:

bash scripts/wan22/inference/distill/run_wan22_moe_i2v_distill.sh

We recommend using the Euler scheduler (default in lightx2v) with the following settings:

• shift=5.0
• guidance_scale=1.0 (i.e., without CFG)

License Agreement

The models in this repository are licensed under the Apache 2.0 License. We claim no rights over the your generate contents, granting you the freedom to use them while ensuring that your usage complies with the provisions of this license. You are fully accountable for your use of the models, which must not involve sharing any content that violates applicable laws, causes harm to individuals or groups, disseminates personal information intended for harm, spreads misinformation, or targets vulnerable populations. For a complete list of restrictions and details regarding your rights, please refer to the full text of the license.

Acknowledgements

We would like to thank the contributors to the Wan2.1, Self-Forcing repositories, for their open research.