Yifan Zhou¹ Zeqi Xiao¹ Shuai Yang² Xingang Pan^1
1S-Lab, Nanyang Technological University,
²Wangxuan Institute of Computer Technology, Peking University

Official PyTorch implementation of Alias-free latent diffusion models.

We found the VAE and denoising network in LDM are not equivariant to fractional shifts. We propose an alias-free framework to improve the fractional shift equivariance of LDM. We demonstrate the effectiveness of our method in various applications, including video editing, frame interpolation, super-resolution and normal estimation.

• Chinese blog posts
• Refine documents
• Training scripts

• [12/2025]: Training code relased.
• [03/2025]: Repository created.

1. Clone the repository. (Don't forget --recursive. Otherwise, please run git submodule update --init --recursive)

git clone [email protected]:SingleZombie/AFLDM.git --recursive
cd AFLDM
pip install -e .

2. Install PyTorch in your Python environment.

3. Install pip libraries.

pip install -r requirements.txt

All the detailed commands are shown inside .sh files.

bash shift_ldm_ffhq.sh

Due to the limitation of our computation resource, the finetuned alias-free Stable Diffusion has a poor generation capacity. It can only perform simple editing.

bash video_editing.sh

bash image_interpolation.sh

This is not a blind SR. The degradation function is fixed.

bash shift_ldm_sr.sh

bash shift_normal_estimation.sh

Download ImageNet (ILSVRC2012_img_train.tar) and extract the sub files. The organization of directory should be like:

train
├── n01440764
└── n01443537
...

1. Update train_data_dir with your ImageNet path in configs/vae/train_afvae_imagenet.json.

2. Run script. bash train_afvae.sh

1. Run script. bash train_afldm.sh

2. Update path in scripts/shift_ldm_ffhq.py with train_ckpt/ffhq_uncond_afldm (the default output diretory set in configs/ldm/train_unet_ffhq.json). Run the script bash shift_ldm_ffhq.sh to test the results.

1. Update train_data_dir with your ImageNet path in configs/sr/train_i2sb_imagenet.json.

2. Run script. bash train_af_i2sb_sr.sh

3. Update path in scripts/shift_ldm_sr.py with train_ckpt/imagenet_sr_i2sb (the default output diretory set in configs/sr/train_i2sb_imagenet.json). Run the script bash shift_ldm_sr.sh to test the results.

@inproceedings{zhou2025afldm,
      title={Alias-Free Latent Diffusion Models: Improving Fractional Shift Equivariance of Diffusion Latent Space},
      author={Zhou, Yifan and Xiao, Zeqi and Yang, Shuai and Pan, Xingang },
      booktitle = {CVPR},
      year = {2025},
    }

• Diffusers: Our project is built on diffusers.
• GMFlow: Our flow estimator.
• StyleGAN3: For sharing alias-free module implementation.
• Alias-Free Convnets: For sharing alias-free module implementation.
• I2SB: For sharing SR implementation.
• StableNormal: For sharing normal estimation dataset.