Awesome-Training-Free-WAN2.1-Editing🍀

Star🌟 is a great help in open source development!

Awesome Training-Free (Inversion-Free) methods meet WAN2.1-T2V.

• WAN2.1 + FlowEdit
• WAN2.1 + FlowAlign
• WANAlign2.1⚡ (FlowAlign + masking method)

News 📑

• FlowAlign (ICLR 26) cited my method in the video editing section of their paper🌠
• [2025-12-16] Introduced on the official FlowEdit GitHub🔥
• [2025-12-14] Introduced on the official WAN GitHub🔥

Introduce WANAlign2.1⚡

I present WANAlign2.1⚡, an inversion-free video editing framework that combines the inversion-free editing method FlowAlign with WAN2.1. By integrating FlowAlign’s inversion-free sampling equation into WAN2.1, our approach preserves the intrinsic characteristics of the source video during editing.
To further enhance control, I introduce Decoupled Inversion-Free Sampling (DIFS), which leverages attention masking to independently adjust the editing strength between preserved and modified regions.
The previous methods frequently modified regions that should have been preserved, thereby degrading overall consistency. The WANAlign2.1⚡ achieves improved spatial-temporal consistency and enhanced text-guided editing performance through DIFS.
As shown in the Results, qualitative results demonstrate that our method is state-of-the-art.

Results🐦‍🔥

0️⃣Inference speed

A100 80GB GPU

1️⃣Color/Background Editing

2️⃣Object Editing

3️⃣Texture Editing

4️⃣Add Effect/Object

Quick Start (Code)🥏

Environment

git clone https://github.com/KyujinHan/Awesome-Training-Free-WAN2.1-Editing.git
cd ./Awesome-Training-Free-WAN2.1-Editing

conda create -n wanalign python=3.10 -y
conda activate wanalign
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.1 -c pytorch -c nvidia

pip install matplotlib omegaconf imageio
pip install transformers==4.51.3 accelerate
pip install imageio[ffmpeg] ftfy

You must install provided diffusers library.

cd ./diffusers
pip install -e .

I used one A100 80GB GPU.

Run code🏂

python awesome_wan_editing.py --config=[__config_yaml_path__]
# python awesome_wan_editing.py --config=./config/object_editing/bear_tiger.yaml

There are some config file examples.

In FlowAlign, there is zeta_scale.
If the value is high, it will be similar to the source video.
The bg_zeta_scale value is only activated when flag_attnmask is True.

Detail Code lines🏫

• FlowEdit Code: WanPipeline.flowedit
• FlowAlign Code: WanPipeline.flowalign
• WANAlign2.1⚡ Core Code: WanPipeline.flowalign
• Attention Extract Code: wan_attention

If you want to visualize attention masking maps, please activate these code.

TODO-list

• Integrating Diffusers🤗
• Code Release
• Support WAN2.1-T2V-1.3B

References

• WAN2.1-T2V-1.3B
• MasaCtrl
• FlowEdit
• FlowAlign
• FlowDirector