Disentangling Disentangled Representations: Towards Improved Latent Units via Diffusion Models

Paper

Prepare data

• Cars3d
  Download nips2015-analogy-data.tar.gz
• Shapes3d
  Download 3dshapes.h5
• MPI3D-toy Download mpi3d_toy.npz
• CelebA Download CelebA/Img (CelebA) or Download celeba-dataset.zip (Kaggle)

Prepare the following datasets in the specified format:

• cars3d
• shapes3d
• mpi3d
• celeba

code/datasets/dataset.py

Place the images in the directory structure:
code/datasets/dataset_name/image_{number}.png

Training

To start the training process, run the following script:

code/scripts/train.sh

Evaluation

To evaluate the trained model, use the following script:

code/scripts/eval.sh

Latent Interchange

For latent interchange operations, run the script:

code/scripts/latent-interchange.sh

Image Generation for FID

To generate images for the Fréchet Inception Distance (FID) score calculation, use:

code/scripts/sampling-for-fid.sh

Files

Cars3d

• First Stage Model: Download (Drive)
• Model: Download (Drive)
• Representations: Download (Drive)
• Gaussians: Download (Drive)
• Label CSV (for Test): Download (Drive)

Shapes3d

• First Stage Model: Download (Drive)
• Model: Download (Drive)
• Representations: Download (Drive)
• Gaussians: Download (Drive)
• Label CSV (for Test): Download (Drive)

MPI3D-toy

• First Stage Model: Download (Drive)
• Model: Download (Drive)
• Representations: Download (Drive)
• Gaussians: Download (Drive)
• Label CSV (for Test): Download (Drive)

CelebA

• First Stage Model: Download (Drive)
• Model: Download (Drive)
• Representations: Download (Drive)
• Gaussians: Download (Drive)
• Label CSV (for Test, included in the dataset): Download (Kaggle)

Acknowledgement

We based our codes on ThomasMrY/DisDiff