/Latent-DDPMs

In this Repository, we test the performance of different backbones on Latent Diffusion task for Image generation.

Primary LanguagePython

Latent-DDPMs

In this Repository, we test the performance of different backbones on Latent Diffusion task for Image generation.

Tested Backbones

  1. UNet architecture proposed in the original Paper.
  2. Wide-ResNet Based Backbone Network
  3. EfficientNetV3 based Backbone Network
  4. VisionTransformer Based Backnoe Network

Implementation Details

  1. Image size : We train on 256x256 sized images, but we use a pretrained network to bring the size of the embeddings down to 1x1x512. The diffusion is carried out in this Latent space.
  2. Schedule : We use a linear noise decay schedule
  3. Dataset : We train the network on : a. MNIST dataset. b. Subset of the WikiArt dataset.

TODO :

  1. Cite the original Paper
  2. Cite datasets
  3. Cite reference code implementations